![\"GA](\"https://preview.ibb.co/faD8kf/ga-history.png\")
UPDATE: See update post in thread.
\n\nWe’ve been made aware that the Google Merchandise Store Demo Account serves as an external source for mining the target variable. Despite our best efforts to have eyes on every corner of potential access to a competition's ground truth, this account’s existence was regrettably unknown to both Kaggle and the host team. And we’re deeply sorry for the impact it’s had on the past month of hard work you’ve put in. Thank you to those (including Jerome) who alerted us to the compromised access to the test set’s solution.
\n\nWe are currently investigating a revision to the original problem which will enable us to keep the competition running and ideally minimize the disruption to your existing work. In the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data. There may be a slight revision to the measurement of the target variable, but we are doing our best to keep in tact the directional intent of the original problem.
\n\nPlease trust that coming to a decision on this and updating you is a top priority for us, and the time we are taking is with the hope that this will be the very final change required for this competition. Further updates with the go forward decision will be posted to this thread.
", "messages": [ { "id": 402520, "postDate": "2018-10-11T20:26:42.743Z", "content": "UPDATE: See update post in thread.
\n\nWe’ve been made aware that the Google Merchandise Store Demo Account serves as an external source for mining the target variable. Despite our best efforts to have eyes on every corner of potential access to a competition's ground truth, this account’s existence was regrettably unknown to both Kaggle and the host team. And we’re deeply sorry for the impact it’s had on the past month of hard work you’ve put in. Thank you to those (including Jerome) who alerted us to the compromised access to the test set’s solution.
\n\nWe are currently investigating a revision to the original problem which will enable us to keep the competition running and ideally minimize the disruption to your existing work. In the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data. There may be a slight revision to the measurement of the target variable, but we are doing our best to keep in tact the directional intent of the original problem.
\n\nPlease trust that coming to a decision on this and updating you is a top priority for us, and the time we are taking is with the hope that this will be the very final change required for this competition. Further updates with the go forward decision will be posted to this thread.
", "rawMarkdown": "**UPDATE: See [update post in thread][1].**\n\nWe’ve been made aware that the [Google Merchandise Store Demo Account][2] serves as an external source for mining the target variable. Despite our best efforts to have eyes on every corner of potential access to a competition's ground truth, this account’s existence was regrettably unknown to both Kaggle and the host team. And we’re deeply sorry for the impact it’s had on the past month of hard work you’ve put in. Thank you to those (including [Jerome][3]) who [alerted us][4] to the compromised access to the test set’s solution. \n\nWe are currently investigating a revision to the original problem which will enable us to keep the competition running and ideally minimize the disruption to your existing work. In the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data. There may be a slight revision to the measurement of the target variable, but we are doing our best to keep in tact the directional intent of the original problem. \n\nPlease trust that coming to a decision on this and updating you is a top priority for us, and the time we are taking is with the hope that this will be the very final change required for this competition. Further updates with the go forward decision will be posted to this thread.\n\n\n [1]: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/68353#404989\n [2]: https://support.google.com/analytics/answer/6367342#access\n [3]: https://www.kaggle.com/emorej\n [4]: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/68235#401880", "votes": 87 }, { "id": 404989, "postDate": "2018-10-16T17:49:22.597Z", "content": "Thank you all for your patience as we chart a way forward. We’ve decided to make the following updates to the competition design.
\n\nWith this design update confirmed, you should be able to continue building your models with the existing data provided. Once the updated datasets are provided, we’ll update this thread and you will have at least 3 weeks time to refine your models with the actual data.
\n\nAgain, thank you for your participation and bearing with us through this challenging competition design update. We’ve done our very best to minimize the disruption to modeling you may already have done, while ensuring the integrity of the competition’s results.
", "rawMarkdown": "Thank you all for your patience as we chart a way forward. We’ve decided to make the following updates to the competition design.\n\n- The competition end date has been extended until November 30, 2018 to compensate for the time required to revise the dataset and design.\n- We will release an updated dataset that extends the timeframe of available data from August 1, 2016 through October 15, 2018. This will be further split between a training set and public leaderboard test set. The public leaderboard will contain a timeframe whose target variable is publicly available via the demo account, to be used as a mechanism for honest submitters to use for cross validation.\n- The test set will be future-looking, composed of all transactions from December 1, 2018 through January 31, 2019. You will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.\n- The [Google Merchandise Store Demo Account][1] effectively becomes freely available to use public data, but will no longer provide any additional benefit to final standings.\n- The train, test & sample submission datasets will be updated in 2-3 weeks time, but no later than November 9, 2018. While the time gap is not ideal, it is required for us to pull and format the data needed.\n- The competition’s private leaderboard will be scored and refreshed on February 8, 2019, at which point the standings will be final.\n- The prize money and allocations will remain in tact, but its distribution will be delayed until after the final standings are confirmed and winners’ obligations are fulfilled.\n\nWith this design update confirmed, you should be able to continue building your models with the existing data provided. Once the updated datasets are provided, we’ll update this thread and you will have at least 3 weeks time to refine your models with the actual data.\n\nAgain, thank you for your participation and bearing with us through this challenging competition design update. We’ve done our very best to minimize the disruption to modeling you may already have done, while ensuring the integrity of the competition’s results.\n\n\n[1]: https://support.google.com/analytics/answer/6367342#access", "votes": 40, "replies": [ { "id": 404994, "postDate": "2018-10-16T17:52:41.757Z", "content": "Thanks for the update and the redesign of competition to make it practical/realistic.
", "rawMarkdown": "Thanks for the update and the redesign of competition to make it practical/realistic." }, { "id": 405001, "postDate": "2018-10-16T18:01:36.357Z", "content": "Finally an update which resolves but doesn't resolve all of major concerns.
\n\n\n\n\nThe public leaderboard will contain a timeframe whose target variable\n is publicly available via the demo account, to be used as a mechanism\n for honest submitters to use for cross validation.
\n
I agree on the honesty part. What about leaderboard probing?
", "rawMarkdown": "Finally an update which resolves but doesn't resolve all of major concerns. \n\n> The public leaderboard will contain a timeframe whose target variable\n> is publicly available via the demo account, to be used as a mechanism\n> for honest submitters to use for cross validation.\n\nI agree on the honesty part. What about leaderboard probing? " }, { "id": 405005, "postDate": "2018-10-16T18:12:20.670Z", "content": "Great idea! Thank you for addressing the issue, kaggle! Now it's a 'real-world' problem! No more LB probing!
", "rawMarkdown": "Great idea! Thank you for addressing the issue, kaggle! Now it's a 'real-world' problem! No more LB probing! " }, { "id": 405011, "postDate": "2018-10-16T18:17:32.210Z", "content": "Thanks... I pity those poor ignorant dishonest competitors ;)!!!!
", "rawMarkdown": "Thanks... I pity those poor ignorant dishonest competitors ;)!!!!", "votes": 4 }, { "id": 405014, "postDate": "2018-10-16T18:22:09.823Z", "content": "Thanks for update - I don't think you needed to add \"honest\" as I believe it to be pejorative and unnecessary. I assume all other external data will be allowed?
", "rawMarkdown": "Thanks for update - I don't think you needed to add \"honest\" as I believe it to be pejorative and unnecessary. I assume all other external data will be allowed?", "votes": 2 }, { "id": 405017, "postDate": "2018-10-16T18:27:26.560Z", "rawMarkdown": "", "isDeleted": true }, { "id": 405019, "postDate": "2018-10-16T18:28:40.357Z", "content": "Thanks for the update and the objective of allowing us to re-use some existing work. Does this mean the competition will be kernels only and you will run our kernels again after 31st January 2019 even though we submitted them by November 30th? Or, does it mean that we need to make the prediction for the test period by November 30th without having the session information for sessions from December 1st?
", "rawMarkdown": "Thanks for the update and the objective of allowing us to re-use some existing work. Does this mean the competition will be kernels only and you will run our kernels again after 31st January 2019 even though we submitted them by November 30th? Or, does it mean that we need to make the prediction for the test period by November 30th without having the session information for sessions from December 1st?", "votes": 25 }, { "id": 405020, "postDate": "2018-10-16T18:29:27.250Z", "content": "nvm
", "rawMarkdown": "nvm", "votes": 1 }, { "id": 405023, "postDate": "2018-10-16T18:36:02.333Z", "content": "@Julia Elliott
\n\nThe Google Merchandise Store Demo Account effectively becomes freely available to use public data, but will no longer provide any additional benefit to final standings.
\n\nIt would be much simpler that Kaggle provides the Merchandise Store Demo instead of making us losing time to find how to get this file. It seems that except Jerome Vallet no one did succeed to find it.
\n\nAs a matter of proof, the second rank today is quite the double of Jerome's score.
\n\nMy advise is : \nNo additional Data OR Available additional Data for everyone.
\n\nThanks a lot for your post !
", "rawMarkdown": "@Julia Elliott\n\n*The Google Merchandise Store Demo Account effectively becomes freely available to use public data, but will no longer provide any additional benefit to final standings.*\n\nIt would be much simpler that Kaggle provides the Merchandise Store Demo instead of making us losing time to find how to get this file. It seems that except Jerome Vallet no one did succeed to find it. \n\nAs a matter of proof, the second rank today is quite the double of Jerome's score. \n\nMy advise is : \nNo additional Data **OR** Available additional Data for everyone.\n\nThanks a lot for your post !\n", "votes": 32 }, { "id": 405035, "postDate": "2018-10-16T19:05:06.207Z", "content": "Will we get visit information for the testing set, or will we only get the IDs?
", "rawMarkdown": "Will we get visit information for the testing set, or will we only get the IDs?", "votes": 1 }, { "id": 405038, "postDate": "2018-10-16T19:07:14.223Z", "content": "We should probably expect huge shakeup on Feb 8
", "rawMarkdown": "We should probably expect huge shakeup on Feb 8", "votes": 3 }, { "id": 405046, "postDate": "2018-10-16T19:18:01.190Z", "content": "Hi Julia, thank you for the updates. Much appreciated! Just a quick question:
\n\nWith the period of the actual private dataset being changed to December 1, 2018 through January 31, 2019, how can Kaggle publish test dataset that is supposed to contain session information such as page views or hits that can only be available in the future?
\n\nThanks,
", "rawMarkdown": "Hi Julia, thank you for the updates. Much appreciated! Just a quick question:\n\nWith the period of the actual private dataset being changed to December 1, 2018 through January 31, 2019, how can Kaggle publish test dataset that is supposed to contain session information such as page views or hits that can only be available in the future? \n\nThanks,\n", "votes": 20 }, { "id": 405049, "postDate": "2018-10-16T19:26:05.360Z", "content": "Please answer @datahobbit and @chriscc questions. This is the most important thing to me, because I don't think we will have information like pageviews, visits, etc. for the future test data. Therefore our validation strategy should change to suppose that we do not know the validation information at all, even if there was a page visit!
", "rawMarkdown": "Please answer @datahobbit and @chriscc questions. This is the most important thing to me, because I don't think we will have information like pageviews, visits, etc. for the future test data. Therefore our validation strategy should change to suppose that we do not know the validation information at all, even if there was a page visit!", "votes": 14 }, { "id": 405051, "postDate": "2018-10-16T19:29:15.957Z", "content": "Julia, thanks for update. \nAs I understand we will use all data currently available, but because we have no future info, we will blindly predict log(1 + sum(revenue)) for each known fullVisitorId. Am I correct?
", "rawMarkdown": "Julia, thanks for update. \nAs I understand we will use all data currently available, but because we have no future info, we will blindly predict log(1 + sum(revenue)) for each known fullVisitorId. Am I correct?" }, { "id": 405068, "postDate": "2018-10-16T20:08:55.730Z", "content": "Please answer @mezoganet query. I guess most of us have the same problem.
\n\nI spent hours working around with the Demo Account finding the right dataset but couldn't.
", "rawMarkdown": "Please answer @mezoganet query. I guess most of us have the same problem. \n\nI spent hours working around with the Demo Account finding the right dataset but couldn't.", "votes": 5 }, { "id": 405070, "postDate": "2018-10-16T20:17:03.983Z", "content": "@Julia - your description leads me to the following remarks.
\n\na. The remaining question is how to combine the competitors predcitions fronm before Nov 30 2018, 23:99 with the data from after Jan 31, 2019, 23:99 . Either we provide the models and Kaggle runs them or Kaggle provides the data and we run them. Or do we just guess before that data even exists? Then this is becoming a lottery.
\n\nb. Your proposal to only score the customers that also were found in the period up to Oct 15 2018 in the data. This profoundly changes the competition goal, because at this moment we did not filter out the non recurring visitors and they were the major part of the test set. The second big diference is that we are now not aiming at 11 months of test data, but only at 2 months that include the Xmas period, which is an outlier period in the current train set. Will we have one or two Xmax periods in the training set? Without much more clarity here we cannot even start thinking about a good modelling strategy...
\n\nc. You use the term \"test set\" in two different way. In your third bullit it is data from Dec 1, 2018 to Jan 31, 2019/ In your fifth bullit it is a a proportions of the data, with unknown start date (yet) but ending Oct 15, 2018. Very inaccurate and confusing.
", "rawMarkdown": "@Julia - your description leads me to the following remarks.\n\na. The remaining question is how to combine the competitors predcitions fronm before Nov 30 2018, 23:99 with the data from after Jan 31, 2019, 23:99 . Either we provide the models and Kaggle runs them or Kaggle provides the data and we run them. Or do we just guess before that data even exists? Then this is becoming a lottery. \n\nb. Your proposal to only score the customers that also were found in the period up to Oct 15 2018 in the data. This profoundly changes the competition goal, because at this moment we did not filter out the non recurring visitors and they were the major part of the test set. The second big diference is that we are now not aiming at 11 months of test data, but only at 2 months that include the Xmas period, which is an outlier period in the current train set. Will we have one or two Xmax periods in the training set? Without much more clarity here we cannot even start thinking about a good modelling strategy...\n\nc. You use the term \"test set\" in two different way. In your third bullit it is data from Dec 1, 2018 to Jan 31, 2019/ In your fifth bullit it is a a proportions of the data, with unknown start date (yet) but ending Oct 15, 2018. Very inaccurate and confusing. \n\n", "votes": 28 }, { "id": 405073, "postDate": "2018-10-16T20:27:15.110Z", "content": "And the plot thickens.... ;)
", "rawMarkdown": "And the plot thickens.... ;)", "votes": 1 }, { "id": 405084, "postDate": "2018-10-16T20:55:52.663Z", "content": "This is getting more confusing. The objective of this competition seems to be like:\n1) Identify if there is a data leakage\n2) If yes, can you provide the best ways to deal with it? p.s. Only word files accepted.
", "rawMarkdown": "This is getting more confusing. The objective of this competition seems to be like:\n1) Identify if there is a data leakage\n2) If yes, can you provide the best ways to deal with it? p.s. Only word files accepted.", "votes": 7 }, { "id": 405091, "postDate": "2018-10-16T21:08:06.940Z", "content": "@MrOoijer
\n\nI have a simple answer to all your complex questions !
\n\nRead the famous Olivier's (@ogrellier) script : I_have_seen_the_future
\n\n:-)
", "rawMarkdown": "@MrOoijer\n\nI have a simple answer to all your complex questions ! \n\nRead the famous Olivier's (@ogrellier) script : **I_have_seen_the_future**\n\n:-)" }, { "id": 405096, "postDate": "2018-10-16T21:14:20.480Z", "content": "@juliaelliott thanks for the update.
\n\nI also agree with @mrooijer that this problem is becoming unpredictable. I appreciate that this is a more real scenario, but now we have to guess if the customer will comeback in the next two months, if he is going to buy anything and how much he will spend only with previous data..
", "rawMarkdown": "@juliaelliott thanks for the update. \n\nI also agree with @mrooijer that this problem is becoming unpredictable. I appreciate that this is a more real scenario, but now we have to guess if the customer will comeback in the next two months, if he is going to buy anything and how much he will spend only with previous data..", "votes": 1 }, { "id": 405108, "postDate": "2018-10-16T21:56:33.170Z", "content": "@datahobbit - good question! It'll be the latter. You'll submit predictions for the users in the test set for that time period (December 1st 2018 to January 31st 2019) by November 30th.
", "rawMarkdown": "@datahobbit - good question! It'll be the latter. You'll submit predictions for the users in the test set for that time period (December 1st 2018 to January 31st 2019) by November 30th.", "votes": 2 }, { "id": 405111, "postDate": "2018-10-16T21:58:44.213Z", "content": "@pfichou - sure! There's a link to the demo account on the Data page. As soon as you sign into the demo account you should be presented with a dashboard allowing you to access Google Merchandise Store data.
", "rawMarkdown": "@pfichou - sure! There's a link to the demo account on the Data page. As soon as you sign into the demo account you should be presented with a dashboard allowing you to access Google Merchandise Store data." }, { "id": 405113, "postDate": "2018-10-16T21:59:42.703Z", "content": "Hey @alexfir - that's correct. You'll need to predict the competition target for unseen data.
", "rawMarkdown": "Hey @alexfir - that's correct. You'll need to predict the competition target for unseen data.", "votes": 1 }, { "id": 405114, "postDate": "2018-10-16T22:01:24.373Z", "content": "@anirudhprakash - hmm, sorry about that. When I signed into the demo account I was immediately at the dashboard for the Google Merchandise Store data. What are you seeing when you log in?
", "rawMarkdown": "@anirudhprakash - hmm, sorry about that. When I signed into the demo account I was immediately at the dashboard for the Google Merchandise Store data. What are you seeing when you log in?" }, { "id": 405115, "postDate": "2018-10-16T22:03:39.173Z", "content": "Hey @chriscc - the competition becomes a problem where you're trying to predict the target for unseen data. So you won't have access to the data for December 1st 2018 to January 31st 2019.
", "rawMarkdown": "Hey @chriscc - the competition becomes a problem where you're trying to predict the target for unseen data. So you won't have access to the data for December 1st 2018 to January 31st 2019.", "votes": 1 }, { "id": 405120, "postDate": "2018-10-16T22:21:12.500Z", "content": "Seems like the Kaggle team figured out a creative way to make this competition possible. I didn't think it was possible, but you've come up with a solution!
\n\nA few comments:
\n\n1) The public leaderboard seems like it will be just for fun, with a huge shake up expected in February! Hopefully most people play along.
\n\n2) The models we have built so far are pretty much useless are they rely on future data such as pageviews/# of visits during the test period.
\n\n3) Will we have to make a different submission for the public LB test period and for the real test period?
\n\n4) People who have extracted data from October 15 to Nov 30 will have an advantage. Hopefully a good samaritan makes it available. We might all have to update our models a few minutes before the deadline!
\n\n5) The new problem seems now to be more realistic, and makes more sense from a business point of view. It will also be much harder to make good predictions!
", "rawMarkdown": "Seems like the Kaggle team figured out a creative way to make this competition possible. I didn't think it was possible, but you've come up with a solution!\n\nA few comments:\n\n1) The public leaderboard seems like it will be just for fun, with a huge shake up expected in February! Hopefully most people play along.\n\n2) The models we have built so far are pretty much useless are they rely on future data such as pageviews/# of visits during the test period. \n\n3) Will we have to make a different submission for the public LB test period and for the real test period? \n\n4) People who have extracted data from October 15 to Nov 30 will have an advantage. Hopefully a good samaritan makes it available. We might all have to update our models a few minutes before the deadline!\n\n5) The new problem seems now to be more realistic, and makes more sense from a business point of view. It will also be much harder to make good predictions!\n\n", "votes": 15 }, { "id": 405124, "postDate": "2018-10-16T22:28:35.893Z", "content": "How are users who don't visit during the test period handled in the scoring process? Completely ignored or is the target set to zero?
", "rawMarkdown": "How are users who don't visit during the test period handled in the scoring process? Completely ignored or is the target set to zero?", "votes": 9 }, { "id": 405129, "postDate": "2018-10-16T22:52:44.447Z", "content": "\n\n\nYou will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.
\n
As Aguiar said, will we predict in this competition the users who will be return and buy something without the detail data in the future, include geo-information, user behavior and traffic source?
", "rawMarkdown": "> You will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.\n\nAs [Aguiar said](https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/68353#405096), will we predict in this competition the users who will be return and buy something without the detail data in the future, include geo-information, user behavior and traffic source?\n\n" }, { "id": 405153, "postDate": "2018-10-17T01:04:29.047Z", "content": "so thats a lottery?
\n\nEdit this was a reply to:
\n\ndatahobbit - good question! It'll be the latter. You'll submit predictions for the users in the test set for that time period (December 1st 2018 to January 31st 2019) by November 30th.
", "rawMarkdown": "so thats a lottery?\n\n Edit this was a reply to:\n\ndatahobbit - good question! It'll be the latter. You'll submit predictions for the users in the test set for that time period (December 1st 2018 to January 31st 2019) by November 30th.", "votes": 5 }, { "id": 405171, "postDate": "2018-10-17T02:03:47.890Z", "content": "Thank you Julia for the update...
\n\nAs I understand correctly from your update and subsequent discussions, we are asked to predict visitor's revenue who visited during test period (all fullVisitorIds in test_v2.csv) i.e., till 15th of Oct'2018. I also guess our submitted model will not be re-run during assessment period. I have following question...
\n\nCustomer buying cycle is a sequence of events i.e., awareness -> Research -> Comparison -> Purchase -> Retention. However, model should predict visitor's 'Purchase' or 'Retention' events without looking customer's behavior?
", "rawMarkdown": "Thank you Julia for the update...\n\nAs I understand correctly from your update and subsequent discussions, we are asked to predict visitor's revenue who visited during test period (all fullVisitorIds in test_v2.csv) i.e., till 15th of Oct'2018. I also guess our submitted model will not be re-run during assessment period. I have following question...\n\n Customer buying cycle is a sequence of events i.e., awareness -> Research -> Comparison -> Purchase -> Retention. However, model should predict visitor's 'Purchase' or 'Retention' events without looking customer's behavior? \n", "votes": 1 }, { "id": 405220, "postDate": "2018-10-17T04:33:17.633Z", "content": "Based on the data description (from Data Page)
\n\n\n\n\ntrain_v2.csv - the updated training set - contains user transactions from August 1st 2016 to April 30th 2018.
\n \ntest_v2.csv - the updated test set - contains user transactions from May 1st 2018 to October 15th 2018.
\n \nsample_submission_v2.csv - a updated sample submission file in the correct format. Contains all fullVisitorIds in test_v2.csv. Your submission's PredictedLogRevenue column should make forward-looking predictions for each of these fullVisitorIds for the timeframe of December 1st 2018 to January 31st 2019. Review \"What am I predicting?\" above for details.
\n
I have following questions:
\n\nWill test_v2.csv contain \"transactionRevenue\" column ? Since we will use data from 2016-08-01 to 2018-10-15, to predict \"transactionRevenue\" from 2018-12-01 to 2019-01-31.
\n\nWe submit predictions for 2018-12-01 to 2019-01-31 ( private LB ). So how is public LB evaluated ? Do we need to submit two predictions ? One for public LB ( test set period, from 2018-05-01 to 2018-10-15 ), one for private LB ( predictions between 2018-12-01 and 2019-1-31) ?
What about data between 2018-10-16 and 2018-11-30 ? We can't benefit from GA demo account data in this period ?
Thanks for your update.
", "rawMarkdown": "Thanks for your update." }, { "id": 405312, "postDate": "2018-10-17T08:10:20.167Z", "content": "So I guess we can take a rest for 2 to 3 weeks till the data updated?
", "rawMarkdown": "So I guess we can take a rest for 2 to 3 weeks till the data updated?", "votes": 1 }, { "id": 405330, "postDate": "2018-10-17T08:53:16.720Z", "content": "@Julia and Kaggle team - big congrats and thank You. What you have essentialy done:
\n\nI think this is very well thought-through. Congrats.
", "rawMarkdown": "@Julia and Kaggle team - big congrats and thank You. What you have essentialy done:\n\n 1. Eliminate the leak\n 2. Make the competition similar to a real-life business problem (no access to the 'future' after the visit, and also no access to visit characteristics like page views or hits or google trends data for test set).\n\nI think this is very well thought-through. Congrats.", "votes": 3 }, { "id": 405333, "postDate": "2018-10-17T08:57:33.533Z", "content": "@SimonChen. It is really confusing indeed. After a second look it seems like:
\n\ntransactionRevenue for test_v2_LB can probably be mined on Google Analytics. If so, the public LB will be pretty meaningless (but does not count for anything anyway).
\n\nBTW: I don't see that X is May 1st, 2018 on the Data Page at the moment. Where did you find that date?
", "rawMarkdown": "@SimonChen. It is really confusing indeed. After a second look it seems like:\n\n- train_v2 will be August 1, 2016 to X\n- test_v2_LB will be X to October 15, 2018\n- data from October 16 - Nov 30 2018 will not be used (but people might benefit from them)\n- the second test set is what counts. Contains the UserIDs of users existing in train_v2 and/or test_v2_LB\n\ntransactionRevenue for test_v2_LB can probably be mined on Google Analytics. If so, the public LB will be pretty meaningless (but does not count for anything anyway).\n\nBTW: I don't see that X is May 1st, 2018 on the Data Page at the moment. Where did you find that date?" }, { "id": 405352, "postDate": "2018-10-17T10:03:01.990Z", "content": "Demo account provides historical data since 1 November 2015. Could you provide supplemental training set for this period? \nAs new task is forecasting Christmas sales volume some additional historical information will be helpful. That will allow to focus on forecasting model rather than scrubbing and cracking demo account.
", "rawMarkdown": "Demo account provides historical data since 1 November 2015. Could you provide supplemental training set for this period? \nAs new task is forecasting Christmas sales volume some additional historical information will be helpful. That will allow to focus on forecasting model rather than scrubbing and cracking demo account.", "votes": 5 }, { "id": 405429, "postDate": "2018-10-17T13:20:03.507Z", "content": "@Erik Bruin, on Data page, you can see the following description:
\n\n\n\n\ntrain_v2.csv - the updated training set - contains user transactions from August 1st 2016 to April 30th 2018.
\n \ntest_v2.csv - the updated test set - contains user transactions from May 1st 2018 to October 15th 2018.
\n
That means X is May 1st
", "rawMarkdown": "@Erik Bruin, on Data page, you can see the following description:\n\n>**train_v2.csv** - the updated training set - contains user transactions from August 1st 2016 to April 30th 2018.\n\n>**test_v2.csv** - the updated test set - contains user transactions from May 1st 2018 to October 15th 2018.\n\nThat means X is May 1st" }, { "id": 405614, "postDate": "2018-10-17T20:11:45.567Z", "content": "Thanks for the update @juliaelliott & Kaggle team - Would the validity timeline for the Google Cloud Platform coupons also be extended till the new Competition end date, since currently they are only valid till Nov 19, 2018?
", "rawMarkdown": "Thanks for the update @juliaelliott & Kaggle team - Would the validity timeline for the Google Cloud Platform coupons also be extended till the new Competition end date, since currently they are only valid till Nov 19, 2018?", "votes": 1 }, { "id": 405618, "postDate": "2018-10-17T20:23:55.877Z", "content": "@Rituparna: I think it was mentioned elsewhere that it won't be the case, unfortunately.
", "rawMarkdown": "@Rituparna: I think it was mentioned elsewhere that it won't be the case, unfortunately." }, { "id": 405622, "postDate": "2018-10-17T20:28:38.273Z", "content": "Thanks for letting me know, Konrad!
", "rawMarkdown": "Thanks for letting me know, Konrad!" }, { "id": 405629, "postDate": "2018-10-17T20:35:03.203Z", "content": "If I understand correctly the new goal is to predict what revenues for certain visitors will be in 9 months time (Train stops in April and predictions have to go through Jan 2019) without knowing sessions ' details from October 15th.
\n\nPeople smart/greedy enough will still be able to grab data from Google Demo Account from May 2018 to Nov 2018...
\n\nAll competitors won't have the same data ?
", "rawMarkdown": "If I understand correctly the new goal is to predict what revenues for certain visitors will be in 9 months time (Train stops in April and predictions have to go through Jan 2019) without knowing sessions ' details from October 15th.\n\nPeople smart/greedy enough will still be able to grab data from Google Demo Account from May 2018 to Nov 2018...\n\nAll competitors won't have the same data ? ", "votes": 7 }, { "id": 405635, "postDate": "2018-10-17T20:39:40.893Z", "content": "\n\nGood point !
", "rawMarkdown": "@olivier \n\nGood point ! " }, { "id": 405911, "postDate": "2018-10-18T10:16:31.870Z", "content": "@ogrellier indeed, this and also the missing data from Nov 2015 to Jul 2016 goes against the declared intention that
\n\n\n", "rawMarkdown": "@ogrellier indeed, this and also the missing data from Nov 2015 to Jul 2016 goes against the declared intention that\n\n> The Google Merchandise Store Demo Account effectively becomes freely available to use public data, but **will no longer provide any additional benefit to final standings**.", "votes": 2 }, { "id": 406006, "postDate": "2018-10-18T13:38:00.333Z", "content": "The Google Merchandise Store Demo Account effectively becomes freely available to use public data, but will no longer provide any additional benefit to final standings.
\n
Thank you for addressing this issue.
", "rawMarkdown": "Thank you for addressing this issue." }, { "id": 406105, "postDate": "2018-10-18T16:36:52.360Z", "content": "Julia & Kaggle team,
\n\nHi @saravji!
\n\n1) I'm glad that people seem to favor the new design. It took some work to figure out what we should do!
\n\n2) Option a is correct.
3) I intend to include the targets in the test set to level the playing field, yep. I consider the public LB to be relatively meaningless already (dependent only on whether people could access the correct parts of the demo account) and we want people to have as much information as possible to solve the new problem formulation.
\n\n4) We did our best to maximize the ability to reuse work - I know that we're still falling short, and I am truly sorry about that. I hope that the new level of realism and learning opportunities help to make up for it.
", "rawMarkdown": "Hi @saravji!\n\n1) I'm glad that people seem to favor the new design. It took some work to figure out what we should do!\n\n2) Option `a` is correct.\n\n3) I intend to include the targets in the test set to level the playing field, yep. I consider the public LB to be relatively meaningless already (dependent only on whether people could access the correct parts of the demo account) and we want people to have as much information as possible to solve the new problem formulation.\n\n4) We did our best to maximize the ability to reuse work - I know that we're still falling short, and I am truly sorry about that. I hope that the new level of realism and learning opportunities help to make up for it.", "votes": 4 }, { "id": 406272, "postDate": "2018-10-19T00:47:13.127Z", "content": "Thanks for the update!
", "rawMarkdown": "Thanks for the update!" }, { "id": 406321, "postDate": "2018-10-19T03:43:02.020Z", "content": "Hi Phil, thanks for the clarifications which are helpful. Just one more question, the updated data description says:
\n\n\n\n\n** Removed Data Fields**
\n \nSome fields were censored to remove target leakage. The major censored fields are listed below.
\n \n1) hits - This row and nested fields are populated for any and all types of hits. Provides a record of all page visits.
\n \n2) customDimensions - This section contains any user-level or session-level custom dimensions that are set for a session. This is a repeated field and has an entry for each dimension that is set.
\n \n3) totals - Multiple sub-columns were removed from the totals field.
\n
It sounds like hits will be removed but I'm not sure if other attributes under totals, e.g. page views will be removed? Statement 1) and 3) seem inconsistent on page views which is under totals.....
", "rawMarkdown": "Hi Phil, thanks for the clarifications which are helpful. Just one more question, the updated data description says:\n\n> ** Removed Data Fields**\n\n> Some fields were censored to remove target leakage. The major censored fields are listed below.\n\n> 1) hits - This row and nested fields are populated for any and all types of hits. Provides a record of all page visits.\n\n> 2) customDimensions - This section contains any user-level or session-level custom dimensions that are set for a session. This is a repeated field and has an entry for each dimension that is set.\n\n> 3) totals - Multiple sub-columns were removed from the totals field.\n\nIt sounds like hits will be removed but I'm not sure if other attributes under totals, e.g. page views will be removed? Statement 1) and 3) seem inconsistent on page views which is under totals.....", "votes": 1 }, { "id": 407232, "postDate": "2018-10-20T17:09:20.763Z", "content": "@philculliton - thanks for your response indicating that we will make predictions by Nov 30th for visitor ID's in the test set for the period Dec 1st 2018 to Jan 31st 2019. I understand that this will be based only on visitors who were existing in the period August 1, 2016 to October 15, 2018 (see below quote from Julia's update). Apologies if this has already been clarified and I missed it, but it seems to me that we don't know whether those visitors will visit again in the period Dec 1st 2018 to Jan 31st 2019. Are you saying that the scoring of the private leaderboard will treat all those that don't visit again during that period as having a zero revenue? Or, are you saying that those that don't visit again from Dec 1st 2018 to Jan 31st 2019 will be excluded from the private leaderboard scoring?
\n\n\n", "rawMarkdown": "@philculliton - thanks for your response indicating that we will make predictions by Nov 30th for visitor ID's in the test set for the period Dec 1st 2018 to Jan 31st 2019. I understand that this will be based only on visitors who were existing in the period August 1, 2016 to October 15, 2018 (see below quote from Julia's update). Apologies if this has already been clarified and I missed it, but it seems to me that we don't know whether those visitors will visit again in the period Dec 1st 2018 to Jan 31st 2019. Are you saying that the scoring of the private leaderboard will treat all those that don't visit again during that period as having a zero revenue? Or, are you saying that those that don't visit again from Dec 1st 2018 to Jan 31st 2019 will be excluded from the private leaderboard scoring?\n\n> The test set will be future-looking, composed of all transactions from December 1, 2018 through January 31, 2019. You will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.", "votes": 7 }, { "id": 407747, "postDate": "2018-10-21T17:47:45.150Z", "content": "The test set will be future-looking, composed of all transactions from December 1, 2018 through January 31, 2019. You will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.
\n
it would have been better if the update was separate thread, most people might not even know that there is an update.
", "rawMarkdown": "it would have been better if the update was separate thread, most people might not even know that there is an update.", "votes": 3 }, { "id": 408584, "postDate": "2018-10-23T05:46:23.547Z", "content": "Thank you very much for update. But what will be meaning to make Google Merchandise Store Demo Account Data available if it will no longer provide any additional benefit to final standings
", "rawMarkdown": "Thank you very much for update. But what will be meaning to make Google Merchandise Store Demo Account Data available if it will no longer provide any additional benefit to final standings", "votes": 1 }, { "id": 408601, "postDate": "2018-10-23T06:15:36.247Z", "content": "It will be better to make a separate thread when the data and competition description are updated. It is hard to find the updates here...
", "rawMarkdown": "It will be better to make a separate thread when the data and competition description are updated. It is hard to find the updates here...", "votes": 2 }, { "id": 410417, "postDate": "2018-10-26T02:59:27.090Z", "content": "@Julia, I see some issues with the new setting, which I detailed in this post. Do you mind to take a look please?
", "rawMarkdown": "@Julia, I see some issues with the new setting, which I detailed in [this post][1]. Do you mind to take a look please?\n\n\n [1]: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/69592", "votes": 1 }, { "id": 411354, "postDate": "2018-10-27T22:47:30.247Z", "content": "First of all, gj at solving the leakage effectivelly. \nAs a matter of fact, I do not understand the reason of a LB in this case. I believe that the test dataset should be included on the training dataset and then make blind predictions for the dates >1/12/2018 for the IDs included on the submission file
", "rawMarkdown": "First of all, gj at solving the leakage effectivelly. \nAs a matter of fact, I do not understand the reason of a LB in this case. I believe that the test dataset should be included on the training dataset and then make blind predictions for the dates >1/12/2018 for the IDs included on the submission file", "votes": 2 } ] }, { "id": 404252, "postDate": "2018-10-15T13:27:55.923Z", "content": "Hi Julia. I know it was because of the weekend, but it's been 4 days and there is still no any official words coming from Kaggle. I expect you to understand how waiting without any information can be annoying for us. Before tomorrow, it would be great to hear some updates like 'we're still working on' or 'we are discussing this and this...' etc.
", "rawMarkdown": "Hi Julia. I know it was because of the weekend, but it's been 4 days and there is still no any official words coming from Kaggle. I expect you to understand how waiting without any information can be annoying for us. Before tomorrow, it would be great to hear some updates like 'we're still working on' or 'we are discussing this and this...' etc. ", "votes": 29, "replies": [ { "id": 404715, "postDate": "2018-10-16T09:02:45.710Z", "content": "Indeed, and I have seen this before from Kaggle team. What's difficult about saying something like \"we will have an update for you on Wednesday or so\". Staying totally silent for 5 days is not good.
", "rawMarkdown": "Indeed, and I have seen this before from Kaggle team. What's difficult about saying something like \"we will have an update for you on Wednesday or so\". Staying totally silent for 5 days is not good.", "votes": 5 } ] }, { "id": 402528, "postDate": "2018-10-11T20:39:48.943Z", "content": "Weekend off!!! Hooray!!
", "rawMarkdown": "Weekend off!!! Hooray!! ", "votes": 21 }, { "id": 402730, "postDate": "2018-10-12T08:05:44.493Z", "content": "One last thing - hopefully when you restart the competition this won't mean Olivier and others lose there kernel medals through no fault of their own. I know Olivier is one gold away from kernels grandmaster status at the minute and it would be tragic if his hard work gets punished.
", "rawMarkdown": "One last thing - hopefully when you restart the competition this won't mean Olivier and others lose there kernel medals through no fault of their own. I know Olivier is one gold away from kernels grandmaster status at the minute and it would be tragic if his hard work gets punished.", "votes": 22, "replies": [ { "id": 402735, "postDate": "2018-10-12T08:12:50.090Z", "content": "Congrats to Olivier anyway.......4 Gold medals on the competition is incredible! I \"only\" have one, but want to keep that one of course ;-). Although I cannot imagine this happening (we got the votes!), maybe Kaggle could decouple the kernels from the competition if things change drastically (convert them to dataset kernels).
\n\nBTW: I am also going on strike. See no point in continuing if optimization might be useless anyway after changes to the competition.......
", "rawMarkdown": "Congrats to Olivier anyway.......4 Gold medals on the competition is incredible! I \"only\" have one, but want to keep that one of course ;-). Although I cannot imagine this happening (we got the votes!), maybe Kaggle could decouple the kernels from the competition if things change drastically (convert them to dataset kernels).\n\nBTW: I am also going on strike. See no point in continuing if optimization might be useless anyway after changes to the competition.......", "votes": 5 }, { "id": 402744, "postDate": "2018-10-12T08:26:35.320Z", "content": "@scirpus, I completely agree with you. However, I feel one can also argue in the same way that he/she worked really hard (either locally and did not submit or submit and was in the top) and his/her hard work was not paid off.
\n\nI guess the optimization the kaggle should plan is to respect hard work of one month of many kagglers and also bring back the competition spirit. So, can there be any middle ground like - kernel medels, kernel points and let's say 0.5X competition points after freezing the lb as of day before yesterday and again relauch the comp with the same spirit as what they are maybe planning to do now.
", "rawMarkdown": "@scirpus, I completely agree with you. However, I feel one can also argue in the same way that he/she worked really hard (either locally and did not submit or submit and was in the top) and his/her hard work was not paid off. \n\nI guess the optimization the kaggle should plan is to respect hard work of one month of many kagglers and also bring back the competition spirit. So, can there be any middle ground like - kernel medels, kernel points and let's say 0.5X competition points after freezing the lb as of day before yesterday and again relauch the comp with the same spirit as what they are maybe planning to do now.\n", "votes": 3 }, { "id": 402753, "postDate": "2018-10-12T08:47:27.363Z", "content": "Any method that freezes the leaderboard is not acceptable. You cannot punish people for following the announced timeline and working locally without frantically submitting everyday.
", "rawMarkdown": "Any method that freezes the leaderboard is not acceptable. You cannot punish people for following the announced timeline and working locally without frantically submitting everyday.", "votes": 10 }, { "id": 402832, "postDate": "2018-10-12T12:29:02.880Z", "content": "Maybe they can just close this competition and then launch a new but similar competition. That might make it easy Olivier and others to retain their kernel medals.
", "rawMarkdown": "Maybe they can just close this competition and then launch a new but similar competition. That might make it easy Olivier and others to retain their kernel medals.", "votes": 1 } ] }, { "id": 402699, "postDate": "2018-10-12T06:45:22.433Z", "content": "I do hope you are at least think of giving Jerome a small token of your appreciation. This will encourage leaks to be found sooner rather than later.
\n\nAlso your nice banner image above clearly shows something that looks like the google store demo ;)
", "rawMarkdown": "I do hope you are at least think of giving Jerome a small token of your appreciation. This will encourage leaks to be found sooner rather than later.\n\nAlso your nice banner image above clearly shows something that looks like the google store demo ;)", "votes": 20, "replies": [ { "id": 402710, "postDate": "2018-10-12T07:23:47.920Z", "content": "@Scirpus thank you for all your comments!\nJulia Elliott and Phil Culliton thanked me personally and the Kaggle team has just given me a Kaggle Swag :)\nThank you all for your kindness!
", "rawMarkdown": "@Scirpus thank you for all your comments!\nJulia Elliott and Phil Culliton thanked me personally and the Kaggle team has just given me a Kaggle Swag :)\nThank you all for your kindness!", "votes": 40 }, { "id": 402713, "postDate": "2018-10-12T07:26:02.463Z", "content": "Brilliant News - you thoroughly deserve it!\nWell done to Kaggle too!
", "rawMarkdown": "Brilliant News - you thoroughly deserve it!\nWell done to Kaggle too!", "votes": 2 } ] }, { "id": 402533, "postDate": "2018-10-11T20:45:01.390Z", "content": "Julia, thanks for the quick update. It's good to know Kaggle is aware and trying to tackle the issue.
\n\nI really hope you'll be able to modify the challenge in a way ground truth data becomes useless but IMHO this would require significant changes in the terms of the competitions and certainly a lot more than a slight change in the metric or/and the target.
\n\nAs far as I'm concerned I don't feel I can go on with my own work on this competition while Kaggle is trying to come up with a contingency plan and act like nothing happened.
", "rawMarkdown": "Julia, thanks for the quick update. It's good to know Kaggle is aware and trying to tackle the issue.\n\nI really hope you'll be able to modify the challenge in a way ground truth data becomes useless but IMHO this would require significant changes in the terms of the competitions and certainly a lot more than a slight change in the metric or/and the target.\n\nAs far as I'm concerned I don't feel I can go on with my own work on this competition while Kaggle is trying to come up with a contingency plan and act like nothing happened. ", "votes": 16, "replies": [ { "id": 402543, "postDate": "2018-10-11T21:16:35.813Z", "content": "The solution is simple: don't allow the use of external data. This is a ML competition not a leak hunting competition.
", "rawMarkdown": "The solution is simple: don't allow the use of external data. This is a ML competition not a leak hunting competition. ", "votes": 14 }, { "id": 402546, "postDate": "2018-10-11T21:20:20.027Z", "content": "It is funny, because this competition originally started as \"No external data allowed\", but the Kagglers convinced them to change this decision, because we wanted to add information like longitude/latitude, GDP of country, timezone, etc.
\n\nNow, we want to change it back to kernels only/no external data allowed.
", "rawMarkdown": "It is funny, because this competition originally started as \"No external data allowed\", but the Kagglers convinced them to change this decision, because we wanted to add information like longitude/latitude, GDP of country, timezone, etc.\n\nNow, we want to change it back to kernels only/no external data allowed.", "votes": 6 }, { "id": 402552, "postDate": "2018-10-11T21:43:38.227Z", "content": "I think there is no way to make sure that people are not using external data (except with kernels only and top 3 teams). You can just add an external feature to one model which will enter in a 20-model ensemble and kaggle will never know... so allowing external data was a good move (in my opinion). The problem is that they used a subset of features/examples from public available data, which also include the target variable
", "rawMarkdown": "I think there is no way to make sure that people are not using external data (except with kernels only and top 3 teams). You can just add an external feature to one model which will enter in a 20-model ensemble and kaggle will never know... so allowing external data was a good move (in my opinion). The problem is that they used a subset of features/examples from public available data, which also include the target variable", "votes": 5 }, { "id": 402555, "postDate": "2018-10-11T21:52:59.087Z", "content": "@jsaguiar It's just a little confusing to know where the line is drawn. For example, in my model I have labeled English speaking countries based on a Wikipedia page of English speaking countries. I'm not importing any external data, but I am referencing external data this way, without doing a merge between train/test and external data source (e.g. I am doing tr_te[Country=='United States']$english_speaking <- 1; tr_te[Country=='Australia']$english_speaking <- 1; etc.)
So....am I using external data?
\n\nYou can extend this to non-toy examples. I don't want to give away other ones so I can preserve a competitive advantage, but you can probably think of other examples where it gets unclear...
\n\nHow do I know what is safe or not? I emailed Phil about this a couple of days ago but haven't received an answer back, and I don't think they will give me an answer because it's too hard of a question to answer/too subjective.
", "rawMarkdown": "@jsaguiar It's just a little confusing to know where the line is drawn. For example, in my model I have labeled English speaking countries based on a Wikipedia page of English speaking countries. I'm not importing any external data, but I am referencing external data this way, without doing a merge between train/test and external data source (e.g. I am doing `tr_te[Country=='United States']$english_speaking <- 1; tr_te[Country=='Australia']$english_speaking <- 1; etc.`)\n\nSo....am I using external data? \n\n1. If yes, I posit that I just \"know\" English speaking countries as general knowledge (e.g. USA, Australia, UK, etc. speaks English).\n2. If not, I posit that I am using external data by referencing an external dataset available online and typing it into my dataset, as opposed to working solely with train/test columns.\n\nYou can extend this to non-toy examples. I don't want to give away other ones so I can preserve a competitive advantage, but you can probably think of other examples where it gets unclear...\n\nHow do I know what is safe or not? I emailed Phil about this a couple of days ago but haven't received an answer back, and I don't think they will give me an answer because it's too hard of a question to answer/too subjective.", "votes": 9 }, { "id": 402571, "postDate": "2018-10-11T23:23:56.220Z", "content": "Exactly, the only way to know if someone used external data is by looking at the code. That's why I think datasets used for competitions should be proprietary (like in Home Credit) or use some method to prevent cheating (two sigma)
", "rawMarkdown": "Exactly, the only way to know if someone used external data is by looking at the code. That's why I think datasets used for competitions should be proprietary (like in Home Credit) or use some method to prevent cheating (two sigma)", "votes": 1 } ] }, { "id": 404739, "postDate": "2018-10-16T10:16:30.560Z", "content": "Kaggle - you are in danger of losing the goodwill of the competitors\nIn this case No News is definitely not Good News\nWhilst we all realise you are working hard behind the scenes there really is no excuse for zero updates
", "rawMarkdown": "Kaggle - you are in danger of losing the goodwill of the competitors\nIn this case No News is definitely not Good News\nWhilst we all realise you are working hard behind the scenes there really is no excuse for zero updates", "votes": 14, "replies": [ { "id": 404930, "postDate": "2018-10-16T15:32:37.710Z", "content": "It has become like a TV series, they leave you with a glitch at the end of the show and then you keep on searching/thinking as to what is the next thing that's going to happen. The only difference is that you are aware of the exact dates there as to when it would be aired which doesn't seem to be the case with Competition considering the updates. It's really sad to see that even after seeing such a keen interest from the participants considering the submissions everyone has made they don't find it worthy for a clear answer.
", "rawMarkdown": "It has become like a TV series, they leave you with a glitch at the end of the show and then you keep on searching/thinking as to what is the next thing that's going to happen. The only difference is that you are aware of the exact dates there as to when it would be aired which doesn't seem to be the case with Competition considering the updates. It's really sad to see that even after seeing such a keen interest from the participants considering the submissions everyone has made they don't find it worthy for a clear answer.", "votes": 7 }, { "id": 405123, "postDate": "2018-10-16T22:27:29.150Z", "content": "Kagglers like @akshay are great at comparing real scenarios and observing things!\nKudos!
", "rawMarkdown": "Kagglers like @akshay are great at comparing real scenarios and observing things!\nKudos!" } ] }, { "id": 403066, "postDate": "2018-10-12T20:11:36.207Z", "content": "I don't know why I keep coming to this thread. Maybe this is the 50th time I am visiting this thread today to check if there is something from Kaggle :(
\n\nMaybe, I need some break.
", "rawMarkdown": "I don't know why I keep coming to this thread. Maybe this is the 50th time I am visiting this thread today to check if there is something from Kaggle :( \n\nMaybe, I need some break. ", "votes": 13, "replies": [ { "id": 403097, "postDate": "2018-10-12T21:14:02.543Z", "content": "", "rawMarkdown": "[You Are Not Alone ][1]\n\n\n [1]: https://www.youtube.com/watch?v=pAyKJAtDNCw", "votes": 11 }, { "id": 403101, "postDate": "2018-10-12T21:32:02.193Z", "content": "Also of interest is how long before other kagglers catch up to Jerome
", "rawMarkdown": "Also of interest is how long before other kagglers catch up to Jerome", "votes": 1 } ] }, { "id": 405374, "postDate": "2018-10-17T11:16:33.787Z", "content": "My main takeaways :
\n\nAll work until now is not completely wasted, but a large part of it is now irrelevant : study of various CV strategies, most of the feature engineering, study of the impact of each feature (which was heavily linked to the way the problem was framed), etc.
The new public leaderboard will also be largely irrelevant. It covers a very different time period from the new private leaderboard (which covers Christmas aka an outlier period). Also the public leaderboard target will be available on the Demo Account, for this reason you'll be able to simulate the public leaderboard locally without being limited to 5 submissions a day. In effect, the public leaderboard becomes yet another possible local CV.
Predicting only based on past history, without the session-level info for the private leaderboard time period, seems like a quite impossible task, considering once again that this is the Christmas period. Someone mentioned a \"lottery\", I'll wait to see what my CV scores are, but I'm not confident it's gonna be any good.
The Demo Account will be a mandatory data source if you have any hope of competing, for at least 3 reasons :
1- Extract the target for the test set/public leaderboard period, in order to not be limited to 5 submissions a day and/or to use it as part of your train set.
\n\n2- Extract data from November 1 2015 to August 1, 2016 to augment the dataset with a 3rd Christmas period.
\n\n3- Extract data from October 15, 2018 to November 30, 2018 to augment the dataset with 15 possibly precious days of data, the closest ones to the private leaderboard period.
\n\nEdit : Persons posting public kernels with the result of these 3 extractions will get an upvote and infinite good karma coming from me at least :)
", "rawMarkdown": "My main takeaways :\n\n* All work until now is not completely wasted, but a large part of it is now irrelevant : study of various CV strategies, most of the feature engineering, study of the impact of each feature (which was heavily linked to the way the problem was framed), etc.\n\n* The new public leaderboard will also be largely irrelevant. It covers a very different time period from the new private leaderboard (which covers Christmas aka an outlier period). Also the public leaderboard target will be available on the Demo Account, for this reason you'll be able to simulate the public leaderboard locally without being limited to 5 submissions a day. In effect, the public leaderboard becomes yet another possible local CV.\n\n* Predicting only based on past history, without the session-level info for the private leaderboard time period, seems like a quite impossible task, considering once again that this is the Christmas period. Someone mentioned a \"lottery\", I'll wait to see what my CV scores are, but I'm not confident it's gonna be any good.\n\n* The Demo Account will be a mandatory data source if you have any hope of competing, for at least 3 reasons :\n\n1- Extract the target for the test set/public leaderboard period, in order to not be limited to 5 submissions a day and/or to use it as part of your train set.\n\n2- Extract data from November 1 2015 to August 1, 2016 to augment the dataset with a 3rd Christmas period.\n\n3- Extract data from October 15, 2018 to November 30, 2018 to augment the dataset with 15 possibly precious days of data, the closest ones to the private leaderboard period.\n\nEdit : Persons posting public kernels with the result of these 3 extractions will get an upvote and infinite good karma coming from me at least :)", "votes": 12, "replies": [ { "id": 405709, "postDate": "2018-10-18T01:04:41.320Z", "content": "@timofeev and @juliencs, can you comment on where this 2015 data is? In the User Explorer, I can't view past Mar 9, 2016. Thanks.
", "rawMarkdown": "@timofeev and @juliencs, can you comment on where this 2015 data is? In the User Explorer, I can't view past Mar 9, 2016. Thanks." }, { "id": 405871, "postDate": "2018-10-18T08:56:09.817Z", "content": "Home -> custom date range on a first chart\n
1- Extract the target for the test set/public leaderboard period, in order to not be limited to 5 submissions a day and/or to use it as part of your train set.\n\n2- Extract data from November 1 2015 to August 1, 2016 to augment the dataset with a 3rd Christmas period.\n\n3- Extract data from October 15, 2018 to November 30, 2018 to augment the dataset with 15 possibly precious days of data, the closest ones to the private leaderboard period.\nExcellent summary, @juliencs.
\n\nSince this competition got a bit more lottery-like, I propose Kaggle goes full throttle on the idea and allow more final submissions if you pay for them. You know, like in a real lottery, the more coupons you buy the more chances you have. All the money then could be used to hire some more people so that next time all the data for a competion will be available on the site, not to be mined by community. This is just me whining :)
", "rawMarkdown": " 1- Extract the target for the test set/public leaderboard period, in order to not be limited to 5 submissions a day and/or to use it as part of your train set.\n \n 2- Extract data from November 1 2015 to August 1, 2016 to augment the dataset with a 3rd Christmas period.\n \n 3- Extract data from October 15, 2018 to November 30, 2018 to augment the dataset with 15 possibly precious days of data, the closest ones to the private leaderboard period.\n\nExcellent summary, @juliencs. \n\nSince this competition got a bit more lottery-like, I propose Kaggle goes full throttle on the idea and allow more final submissions if you pay for them. You know, like in a real lottery, the more coupons you buy the more chances you have. All the money then could be used to hire some more people so that next time all the data for a competion will be available on the site, not to be mined by community. This is just me whining :)", "votes": 3 }, { "id": 406098, "postDate": "2018-10-18T16:23:18.643Z", "content": "@juliencs those with demo account data also have access to additional features that were removed in the competition dataset
", "rawMarkdown": "@juliencs those with demo account data also have access to additional features that were removed in the competition dataset " } ] }, { "id": 403557, "postDate": "2018-10-13T23:51:17.903Z", "content": "Since this is not the first time when a leak occurs, maybe it makes sense to introduce a new type of competition called \"reveal hidden leaks\" :)
", "rawMarkdown": "Since this is not the first time when a leak occurs, maybe it makes sense to introduce a new type of competition called \"reveal hidden leaks\" :) ", "votes": 11 }, { "id": 406365, "postDate": "2018-10-19T06:09:43.330Z", "content": "I don't see much point in removing variables due to target leakage now, we're literally predicting the future in the redesigned competition. Excluding them would reward those who go and get them on their own.
", "rawMarkdown": "I don't see much point in removing variables due to target leakage now, we're literally predicting the future in the redesigned competition. Excluding them would reward those who go and get them on their own.", "votes": 9, "replies": [ { "id": 406454, "postDate": "2018-10-19T09:21:07.807Z", "content": "Do you mean stuff like pageviews and hits for the public LB period?
", "rawMarkdown": "Do you mean stuff like pageviews and hits for the public LB period?" }, { "id": 406651, "postDate": "2018-10-19T15:23:51.173Z", "content": "I meant the fields referred to in the \"** Removed Data Fields**\" section. It made sense when this competition was peeking into the future, but not now.
\n\nHow you treat the Public LB is a trickier question; you don't want to make it completely uncorrelated to the Private LB, but on the other hand, the data for the Public LB period would be useful for training the models. If some people can get it anyway, then maybe we should just make it available to everyone and let those who care get a perfect score on the Public LB.
", "rawMarkdown": "I meant the fields referred to in the \"** Removed Data Fields**\" section. It made sense when this competition was peeking into the future, but not now. \n\nHow you treat the Public LB is a trickier question; you don't want to make it completely uncorrelated to the Private LB, but on the other hand, the data for the Public LB period would be useful for training the models. If some people can get it anyway, then maybe we should just make it available to everyone and let those who care get a perfect score on the Public LB.", "votes": 1 }, { "id": 406666, "postDate": "2018-10-19T15:45:58.610Z", "content": "I completely agree.
\n\nI'll wait to hear from @juliaelliott or @philculliton about the exact nature of the v2 dataset. It's clear by now that in this new competition, the \"removed data fields\", the \"removed date period\" and the \"target for the whole train+test period\" are pretty much mandatory in order to compete seriously. I'm hoping to spend my time practicing ML, not learning to code bots scraping HTML data. If it turns out the new dataset is not comprehensive and you need to do the demo account scraping by yourself, I'll give up on this competition and look at the other ones available.
\n\nHopefully we can get this information (not even the dataset, just the information) sooner rather than later.
", "rawMarkdown": "I completely agree.\n\nI'll wait to hear from @juliaelliott or @philculliton about the exact nature of the v2 dataset. It's clear by now that in this new competition, the \"removed data fields\", the \"removed date period\" and the \"target for the whole train+test period\" are pretty much mandatory in order to compete seriously. I'm hoping to spend my time practicing ML, not learning to code bots scraping HTML data. If it turns out the new dataset is not comprehensive and you need to do the demo account scraping by yourself, I'll give up on this competition and look at the other ones available.\n\nHopefully we can get this information (not even the dataset, just the information) sooner rather than later.", "votes": 5 }, { "id": 406760, "postDate": "2018-10-19T20:00:43.890Z", "content": "Hi @dmitriyguller and @juliencs -
\n\nI agree - I plan to include all of the previously removed data fields in the v2 dataset.
", "rawMarkdown": "Hi @dmitriyguller and @juliencs - \n\nI agree - I plan to include all of the previously removed data fields in the v2 dataset.", "votes": 9 }, { "id": 406931, "postDate": "2018-10-20T03:37:57.973Z", "content": "That's awesome @philculliton! I strongly support this decision. \nWill you be able to release a small dataset in the expected final structure?\nWe will have only 3 weeks with the new dataset until final deadline, to practically start from scratch with new data, and a new goal, almost none of what was done up to now is relevant, that's a very short time.\nProlonging the competition or releasing part of the data in the final structure would make it easier to come up with a good solution, especially for those of us with less time to do kaggling.
", "rawMarkdown": "That's awesome @philculliton! I strongly support this decision. \nWill you be able to release a small dataset in the expected final structure?\nWe will have only 3 weeks with the new dataset until final deadline, to practically start from scratch with new data, and a new goal, almost none of what was done up to now is relevant, that's a very short time.\nProlonging the competition or releasing part of the data in the final structure would make it easier to come up with a good solution, especially for those of us with less time to do kaggling. ", "votes": 2 }, { "id": 407019, "postDate": "2018-10-20T07:52:59.137Z", "content": "I totally agree with Amir. Although I appreciate the new ideas I will pass if this ends up like a mad scramble for three weeks. Why not add another month? That would also take out the Christmas period.
", "rawMarkdown": "I totally agree with Amir. Although I appreciate the new ideas I will pass if this ends up like a mad scramble for three weeks. Why not add another month? That would also take out the Christmas period.", "votes": 1 } ] }, { "id": 403359, "postDate": "2018-10-13T13:05:59.820Z", "content": "This competition should be closed as it does not make any more sense to find or exploit already existing leaks simply for the purpose of improving the rank.
", "rawMarkdown": "This competition should be closed as it does not make any more sense to find or exploit already existing leaks simply for the purpose of improving the rank.", "votes": 9 }, { "id": 404187, "postDate": "2018-10-15T12:12:35.160Z", "content": "I am just hoping they don't rush to such an extent that they swap one leak for another.
", "rawMarkdown": "I am just hoping they don't rush to such an extent that they swap one leak for another. ", "votes": 7 }, { "id": 402720, "postDate": "2018-10-12T07:43:28.687Z", "content": "Thanks for the update.
\n\n\n\n\nIn the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data.
\n
It is hard to stay motivated and keep working on it with this uncertainty. I feel like our efforts are already wasted.
", "rawMarkdown": "Thanks for the update.\n\n> In the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data.\n\nIt is hard to stay motivated and keep working on it with this uncertainty. I feel like our efforts are already wasted.", "votes": 8, "replies": [ { "id": 402792, "postDate": "2018-10-12T10:23:42.867Z", "content": "Exactly. Until rules are clarified I stop my effort here.
", "rawMarkdown": "Exactly. Until rules are clarified I stop my effort here.", "votes": 3 }, { "id": 402827, "postDate": "2018-10-12T12:09:57.743Z", "content": "This sentence about proceeding is meaningless....there is no other way, old data must be switched with properly obfuscated new data.
", "rawMarkdown": "This sentence about proceeding is meaningless....there is no other way, old data must be switched with properly obfuscated new data.", "votes": 1 } ] }, { "id": 402630, "postDate": "2018-10-12T03:02:00.380Z", "content": "Lots of opinions, so I might as well add to it. I find the whole thing to be amusing and hope we can all keep our sense of humor front and center. Is it true that the 'host team' didn't know of the demo account? The whole purpose of the demo account is to demonstrate the value of Google Analytics! Plus. Erik Bruin came out early with his excellent kernel pointing to the demo account, and I made a direct reference to it in my flow diagram kernel.
\n\nI found the demo account in the second week of the competition, but silly me - the API was locked out and I assumed there was no value to the web interface :( Perhaps I am misunderstanding something - very possible - but it sounds like the people wanting us to analyze Google Analytics data didn't know about the Google Analytics demo. Truth is indeed stranger than fiction!
", "rawMarkdown": "Lots of opinions, so I might as well add to it. I find the whole thing to be amusing and hope we can all keep our sense of humor front and center. Is it true that the 'host team' didn't know of the demo account? The whole purpose of the demo account is to demonstrate the value of Google Analytics! Plus. Erik Bruin came out early with his [excellent kernel](https://www.kaggle.com/erikbruin/google-analytics-eda-lightgbm-screenshots) pointing to the demo account, and I made a direct reference to it in my [flow diagram kernel](https://www.kaggle.com/jpmiller/go-with-the-flow-lb-1-447). \n\nI found the demo account in the second week of the competition, but silly me - the API was locked out and I assumed there was no value to the web interface :( Perhaps I am misunderstanding something - very possible - but it sounds like the people wanting us to analyze Google Analytics data didn't know about the Google Analytics demo. Truth is indeed stranger than fiction!", "votes": 8, "replies": [ { "id": 402659, "postDate": "2018-10-12T04:33:11.337Z", "content": "A lot strange thing could happen in large organization. I have worked with some clients, who did not know how to access a lot of specific information in their own database and need help.
\n\nSome participants guess this competition have other purposes. If the guess was true, Kaggle is taking advantage of us and destructing its own credibility for other company. In my opinion, it is unlikely.
", "rawMarkdown": "A lot strange thing could happen in large organization. I have worked with some clients, who did not know how to access a lot of specific information in their own database and need help. \n\nSome participants guess this competition have other purposes. If the guess was true, Kaggle is taking advantage of us and destructing its own credibility for other company. In my opinion, it is unlikely. ", "votes": 1 }, { "id": 402680, "postDate": "2018-10-12T05:40:40.213Z", "content": "I think RStudio wants people to use R as opposed to python. It is fair enough that they don't know about the demo data. I think also it is fair enough that kaggle didn't know about the data. Then there is Google with various departments and products cloud, bigquery, analytics, store. I think the relevant managers at Google should take responsibility. Darryl
", "rawMarkdown": "I think RStudio wants people to use R as opposed to python. It is fair enough that they don't know about the demo data. I think also it is fair enough that kaggle didn't know about the data. Then there is Google with various departments and products cloud, bigquery, analytics, store. I think the relevant managers at Google should take responsibility. Darryl ", "votes": 3 }, { "id": 402687, "postDate": "2018-10-12T06:05:46.067Z", "content": "Thanks for mentioning my kernel John! At, let's say version 2 of my kernel, I already became suspicious to be honest......when I replaced the initial screenshots with screenshots that matched the date range of train+test. I thought: could it, would it......? I eventually decided to not look into it. Also did not really have the time to dig into it. Maybe I should have raised my concerns earlier......
", "rawMarkdown": "Thanks for mentioning my kernel John! At, let's say version 2 of my kernel, I already became suspicious to be honest......when I replaced the initial screenshots with screenshots that matched the date range of train+test. I thought: could it, would it......? I eventually decided to not look into it. Also did not really have the time to dig into it. Maybe I should have raised my concerns earlier......", "votes": 5 } ] }, { "id": 404585, "postDate": "2018-10-16T02:39:09.713Z", "content": "It's been 4 days already. We need some notice and wish to hear from you, julia.
", "rawMarkdown": "It's been 4 days already. We need some notice and wish to hear from you, julia.", "votes": 5 }, { "id": 405699, "postDate": "2018-10-18T00:28:45.597Z", "content": "By any chance, would anyone (@emorej or @juliaelliott) be willing to provide 1) instructions to get this new REQUIRED data source and 2) provide a dump of the data available up to this month (perhaps Kaggle Datasets or an open git repo)? As I understand, there are 3 years worth, about 30k users, and 4M rows (not sure where I read this; 1 row is day-user-visit summary) and we can only download in batches of 5k from the G-Analytics demo? Further it is not being provided directly by Kaggle (because it will keep getting updated). Somewhere in this tool is the full non aggregated data. Perhaps Kaggle/Google can guide us to siphoning off data to actually make this competition fair?
\n\nEdit: Looks like Mar 9, 2016-present (Where is that 2015 data?). 367K users and up near 1K sessions for some users and up to 128 transactions. For someone really wanting to hunt and download this data to win, they will create a bot that spams Google servers until they get the payload to win the competition with more fine grain info. As an example, I picked out info that I find useful to understand the user and the likeliness of a user spending a certain amount. Although we may not get this in the test set, I could build a model that estimates what will be bought and how many as additional features for the test data if I have this [proposed] training data. It's obviously available, but not readily accessible.
\n\nPurchased Google Bluetooth Headphones for $6,248.75\nAction type\nPAYMENT\nTransaction Id\n30755\nTransaction revenue\n$6,248.75\nTransaction tax\n$437.84\nTransaction shipping\n$6.00\nProduct count\n1\nProduct code\nGGOEGEVB070499\nProduct name\nGoogle Bluetooth Headphones\nProduct revenue\n$6,248.75\nProduct quantity\n125
\n\nEdit 2: After further investigation, I see that the whole G suite of tools provides an impressive set of features. I'm a bit lost in the weeds, but I see that there are 525 fields available. Even latitude/longitude and other \"not in demo\" fields are plainly available now. Most appear to be aggregates, but the cool part (if everyone has fair access) are features like \"avg session duration\" and \"avg time on page\". If only I could build models in the Google suite.
", "rawMarkdown": "By any chance, would anyone (@emorej or @juliaelliott) be willing to provide 1) instructions to get this new REQUIRED data source and 2) provide a dump of the data available up to this month (perhaps Kaggle Datasets or an open git repo)? As I understand, there are 3 years worth, about 30k users, and 4M rows (not sure where I read this; 1 row is day-user-visit summary) and we can only download in batches of 5k from the G-Analytics demo? Further it is not being provided directly by Kaggle (because it will keep getting updated). Somewhere in this tool is the full non aggregated data. Perhaps Kaggle/Google can guide us to siphoning off data to actually make this competition fair?\n\nEdit: Looks like Mar 9, 2016-present (Where is that 2015 data?). 367K users and up near 1K sessions for some users and up to 128 transactions. For someone really wanting to hunt and download this data to win, they will create a bot that spams Google servers until they get the payload to win the competition with more fine grain info. As an example, I picked out info that I find useful to understand the user and the likeliness of a user spending a certain amount. Although we may not get this in the test set, I could build a model that estimates what will be bought and how many as additional features for the test data if I have this [proposed] training data. It's obviously available, but not readily accessible.\n\n\nPurchased Google Bluetooth Headphones for $6,248.75\nAction type\nPAYMENT\nTransaction Id\n30755\nTransaction revenue\n$6,248.75\nTransaction tax\n$437.84\nTransaction shipping\n$6.00\nProduct count\n1\nProduct code\nGGOEGEVB070499\nProduct name\nGoogle Bluetooth Headphones\nProduct revenue\n$6,248.75\nProduct quantity\n125\n\nEdit 2: After further investigation, I see that the whole G suite of tools provides an impressive set of features. I'm a bit lost in the weeds, but I see that there are 525 fields available. Even latitude/longitude and other \"not in demo\" fields are plainly available now. Most appear to be aggregates, but the cool part (if everyone has fair access) are features like \"avg session duration\" and \"avg time on page\". If only I could build models in the Google suite.", "votes": 6, "replies": [ { "id": 405719, "postDate": "2018-10-18T02:01:28.657Z", "content": "You have a good point in that the extra data (e.g. what the customer bought) could make a difference!
", "rawMarkdown": "You have a good point in that the extra data (e.g. what the customer bought) could make a difference!" } ] }, { "id": 404684, "postDate": "2018-10-16T07:51:21.377Z", "content": "Julia please, could you give us new updates!\nWe don't know should we work at the current competition or we need to try something else...
", "rawMarkdown": "Julia please, could you give us new updates!\nWe don't know should we work at the current competition or we need to try something else...", "votes": 6 }, { "id": 404509, "postDate": "2018-10-15T20:51:29.873Z", "content": "I keep coming back to this thread hoping for an update. Really confused about which solution to work on. It would be great if we get an update soon.
", "rawMarkdown": "I keep coming back to this thread hoping for an update. Really confused about which solution to work on. It would be great if we get an update soon.", "votes": 6 }, { "id": 405057, "postDate": "2018-10-16T19:40:52.197Z", "content": "First of all, thanks for keep this competition going. \nOne (pure hypothetical) thought: how much money will someone's need to do some real buys in GStore between (December 1, 2018) through (January 31, 2019) and use this information to win first place in final leaderboad? Do not predict it, do it (well, some part of it)!
", "rawMarkdown": "First of all, thanks for keep this competition going. \nOne (pure hypothetical) thought: how much money will someone's need to do some real buys in GStore between (December 1, 2018) through (January 31, 2019) and use this information to win first place in final leaderboad? Do not predict it, do it (well, some part of it)!", "votes": 3, "replies": [ { "id": 405122, "postDate": "2018-10-16T22:24:38.883Z", "content": "You would need to have had a session ID in the train period. I do, but unfortunately I might have cleared my cookies recently!
", "rawMarkdown": "You would need to have had a session ID in the train period. I do, but unfortunately I might have cleared my cookies recently!", "votes": 1 } ] }, { "id": 404791, "postDate": "2018-10-16T11:47:44.560Z", "content": "Its been 4 days. Is there any update ?
", "rawMarkdown": "Its been 4 days. Is there any update ?", "votes": 3 }, { "id": 403401, "postDate": "2018-10-13T14:48:37.993Z", "content": "I believe many kagglers are still working hard on their solution without noticing this update...
", "rawMarkdown": "I believe many kagglers are still working hard on their solution without noticing this update... ", "votes": 4 }, { "id": 402531, "postDate": "2018-10-11T20:43:23.587Z", "content": "Thanks @Julia.
\n\n\"you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data\"
\n\nThis means that there's no need to put on the \"Leak Hunter\" hat!\nThat's really encouraging, I hate that hat...
", "rawMarkdown": "Thanks @Julia.\n\n\"you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data\"\n\nThis means that there's no need to put on the \"Leak Hunter\" hat!\nThat's really encouraging, I hate that hat...", "votes": 4 }, { "id": 404420, "postDate": "2018-10-15T17:11:06.120Z", "content": "Will you guys extend the deadline because of the data leak issue?
", "rawMarkdown": "Will you guys extend the deadline because of the data leak issue? ", "votes": 2 }, { "id": 402557, "postDate": "2018-10-11T21:57:47.967Z", "content": "I don't know how you would expect people to just keep working on the current problem when we know the problem will change and have no indication of the type of change to expect
", "rawMarkdown": "I don't know how you would expect people to just keep working on the current problem when we know the problem will change and have no indication of the type of change to expect", "votes": 3, "replies": [ { "id": 402575, "postDate": "2018-10-11T23:47:40.893Z", "content": "because plenty of people were far from the honest top submissions and still have much progress to be made. I do agree with the concern and uncertainty though.
", "rawMarkdown": "because plenty of people were far from the honest top submissions and still have much progress to be made. I do agree with the concern and uncertainty though.", "votes": 2 } ] }, { "id": 418317, "postDate": "2018-11-09T16:33:28.897Z", "content": "Guys, just in case you wait some kind of notification in this thread about new data availability, it is already here.\nDo not know why organizers did not post update here.
", "rawMarkdown": "Guys, just in case you wait some kind of notification in this thread about new data availability, it is already here.\nDo not know why organizers did not post update here.", "votes": 1 }, { "id": 405922, "postDate": "2018-10-18T10:41:56.627Z", "content": "Will you reset the LB?
", "rawMarkdown": "Will you reset the LB? ", "votes": 1 }, { "id": 405617, "postDate": "2018-10-17T20:21:58.207Z", "content": "I have to say, based on the description in the \"update\" post, that it might get interesting: the 6 weeks gap between the end of training data and the start of real test raises the difficulty level considerably (unless we all refit our models in the last day and Kaggle servers survive literally everyone having to submit at the last moment).
", "rawMarkdown": "I have to say, based on the description in the \"update\" post, that it might get interesting: the 6 weeks gap between the end of training data and the start of real test raises the difficulty level considerably (unless we all refit our models in the last day and Kaggle servers survive literally everyone having to submit at the last moment). ", "votes": 1 }, { "id": 404164, "postDate": "2018-10-15T11:07:07.480Z", "content": "I had already started working on it. The main gist is how we contribute to thought process. Data Leakage will not contribute much to Data Science.
", "rawMarkdown": "I had already started working on it. The main gist is how we contribute to thought process. Data Leakage will not contribute much to Data Science.", "votes": 1 }, { "id": 402712, "postDate": "2018-10-12T07:24:59.510Z", "content": "Thank you, Julia!\nI hope you will find a good decision, but i think you should public demo account data anyway. I hope it will be ml contest raither than scrapper contes.
", "rawMarkdown": "Thank you, Julia!\nI hope you will find a good decision, but i think you should public demo account data anyway. I hope it will be ml contest raither than scrapper contes.", "votes": 1 }, { "id": 402598, "postDate": "2018-10-12T01:17:52.920Z", "content": "This is a leak we already know. What I am worried about is whether there will be other similar leaks like Google Merchandise Store Demo Account ..can get data
", "rawMarkdown": "This is a leak we already know. What I am worried about is whether there will be other similar leaks like [Google Merchandise Store Demo Account][1] ..can get data \n\n\n [1]: https://support.google.com/analytics/answer/6367342#access", "votes": 1 }, { "id": 402550, "postDate": "2018-10-11T21:33:55.553Z", "content": "Hope that time for this competition will add, if dataset changes)
", "rawMarkdown": "Hope that time for this competition will add, if dataset changes)", "votes": 1 }, { "id": 417661, "postDate": "2018-11-08T15:43:40.497Z", "content": "I hope we'll hear from Kaggle with an update soon, but I also hope it's not going to be rushed. If it takes longer than expected, I'd prefer waiting a little more than participating in another flawed competition.
", "rawMarkdown": "I hope we'll hear from Kaggle with an update soon, but I also hope it's not going to be rushed. If it takes longer than expected, I'd prefer waiting a little more than participating in another flawed competition.", "votes": 2, "replies": [ { "id": 417675, "postDate": "2018-11-08T16:06:54.447Z", "content": "Totally agree - slow and steady wins the race!
", "rawMarkdown": "Totally agree - slow and steady wins the race!", "votes": 1 } ] }, { "id": 417194, "postDate": "2018-11-07T23:20:48.340Z", "content": "Will test_v2.csv contain \"transactionRevenue\" column ?
", "rawMarkdown": "Will test_v2.csv contain \"transactionRevenue\" column ?", "votes": 2 }, { "id": 415764, "postDate": "2018-11-05T16:51:38.640Z", "content": "5 days remaining, no news about the expected update
", "rawMarkdown": "5 days remaining, no news about the expected update", "votes": 2 }, { "id": 406083, "postDate": "2018-10-18T16:06:42.073Z", "content": "Are you going to make sure the test data will be two months only so as to mimic (December 1, 2018) through (January 31, 2019)? We don't want people optimizing models that have vastly more data in test than that of the blind test data. As we are trying to predict the sum of revenues this will reduce the probability of competitors vastly over estimating the revenue if they use 1 model to predict individual entries followed by another predicting the revenue sum. (See olivier's script)
", "rawMarkdown": "Are you going to make sure the test data will be two months only so as to mimic (December 1, 2018) through (January 31, 2019)? We don't want people optimizing models that have vastly more data in test than that of the blind test data. As we are trying to predict the sum of revenues this will reduce the probability of competitors vastly over estimating the revenue if they use 1 model to predict individual entries followed by another predicting the revenue sum. (See olivier's script)", "votes": 2 }, { "id": 405217, "postDate": "2018-10-17T04:28:14.977Z", "content": "Dear @juliaelliott,\nthis is my first ever kaggle competition. I worked on it the last two weekends and have learned a lot.\nBut now I returned to the competition page and almost accidentally find out that everything has changed: in the time since I accepted the competition rules, you have changed the rules, you have changed the deadlines AND you have changed the data.
\n\nYou have the email addresses of everyone who has entered the competition -- why do you not notify people about the new rules and data??? I expect the majority of the participants out there is not even aware of the new situation!
\n\nI do not consider the time I used on this problem wasted, since it contributed to my learning; but regarding the competition it actually is wasted, and I find the way kaggle treats the participants very demotivating.
", "rawMarkdown": "Dear @juliaelliott,\nthis is my first ever kaggle competition. I worked on it the last two weekends and have learned a lot.\nBut now I returned to the competition page and almost accidentally find out that everything has changed: in the time since I accepted the competition rules, you have changed the rules, you have changed the deadlines AND you have changed the data.\n\nYou have the email addresses of everyone who has entered the competition -- why do you not notify people about the new rules and data??? I expect the majority of the participants out there is not even aware of the new situation!\n\nI do not consider the time I used on this problem wasted, since it contributed to my learning; but regarding the competition it actually is wasted, and I find the way kaggle treats the participants very demotivating. ", "votes": 2, "replies": [ { "id": 405253, "postDate": "2018-10-17T05:50:43.257Z", "content": "The rule changes in order to prevent people from taking advantage of the data leakage. You can read the related threads to know more.
\n\nMany participants, myself included, are affected by the changes. But I regard it as a better outcome than everyone exploiting the leaks to build sth useless
", "rawMarkdown": "The rule changes in order to prevent people from taking advantage of the data leakage. You can read the related threads to know more. \n\nMany participants, myself included, are affected by the changes. But I regard it as a better outcome than everyone exploiting the leaks to build sth useless", "votes": 1 }, { "id": 405394, "postDate": "2018-10-17T12:09:43.747Z", "content": "@LaPe
\n\nFirst of all, as a first time Kaggler, your time is not wasted. More the time you deal with these kinds of business conflicts, it will give an idea of how to deal, if such things happen at the workplace.
\n\nYou can learn many techniques from others' kernel and keep learning from their approaches (I did a lot and still learning). Who knows, it might help in another competition.
\n\nIf you are here to just make money by winning in the competition, I won't have any words to say.
\n\nThe Kaggle team is not playing hide and seek game here. They are also working hard to meet the standards without giving the advantage for a few Kagglers. Everyone needs a fair play, right?
\n\nRunning the competition is ten times harder than participating. Trust me.
\n\n(This just my opinion. If any of my words hurt you, please forgive me)
", "rawMarkdown": "@LaPe\n\nFirst of all, as a first time Kaggler, your time is not wasted. More the time you deal with these kinds of business conflicts, it will give an idea of how to deal, if such things happen at the workplace. \n\nYou can learn many techniques from others' kernel and keep learning from their approaches (I did a lot and still learning). Who knows, it might help in another competition. \n\nIf you are here to just make money by winning in the competition, I won't have any words to say.\n\nThe Kaggle team is not playing hide and seek game here. They are also working hard to meet the standards without giving the advantage for a few Kagglers. Everyone needs a fair play, right?\n\nRunning the competition is ten times harder than participating. Trust me. \n\n(This just my opinion. If any of my words hurt you, please forgive me)", "votes": 1 }, { "id": 406213, "postDate": "2018-10-18T21:24:50.967Z", "content": "Dear @rglimitless and @rajacsp,
\n\nI perfectly understand that kaggle has to make changes to the competition, and even at the risk of repeating myself: I don't consider my time wasted. If it was not clear from my post: I am critizing the lack of information and communication. Without following in the discussion threads one would not have become aware of the changes in the competition.
", "rawMarkdown": "Dear @rglimitless and @rajacsp,\n\nI perfectly understand that kaggle has to make changes to the competition, and even at the risk of repeating myself: I don't consider my time wasted. If it was not clear from my post: I am critizing the lack of information and communication. Without following in the discussion threads one would not have become aware of the changes in the competition.", "votes": 1 } ] }, { "id": 405063, "postDate": "2018-10-16T19:49:29.753Z", "content": "I recall someone mentioning that not all transactions were reflected in the modeling dataset, only the latest transaction for the session. Is that the case, and if it is, will it be fixed in the new dataset?
", "rawMarkdown": "I recall someone mentioning that not all transactions were reflected in the modeling dataset, only the latest transaction for the session. Is that the case, and if it is, will it be fixed in the new dataset?", "votes": 2 }, { "id": 404780, "postDate": "2018-10-16T11:32:14.233Z", "content": "Hi Julia,\nPlease inform us ASAP because the saying is right: \"A bad decision is better than no decision\". This 'no decision' situation is frustrating all of us.
", "rawMarkdown": "Hi Julia,\nPlease inform us ASAP because the saying is right: \"A bad decision is better than no decision\". This 'no decision' situation is frustrating all of us.", "votes": 2 }, { "id": 404737, "postDate": "2018-10-16T10:08:34.567Z", "content": "\" In the near future \" usually means A few days or weeks ? I think it's better to tell us \"updating in about a week\" or something like that. Give us a more specific time information, then we'll be more content with other thing.
", "rawMarkdown": " \" In the near future \" usually means A few days or weeks ? I think it's better to tell us \"updating in about a week\" or something like that. Give us a more specific time information, then we'll be more content with other thing.", "votes": 2, "replies": [ { "id": 404796, "postDate": "2018-10-16T11:52:41.743Z", "content": "\"In the near future\", meaning 1 day after merger to make it fun..
", "rawMarkdown": "\"In the near future\", meaning 1 day after merger to make it fun..", "votes": 1 }, { "id": 404970, "postDate": "2018-10-16T17:12:47.353Z", "content": "@SimonChen
\n\n\" In the near future \"
\n\nHem, it might depend on the metric.
", "rawMarkdown": "@SimonChen\n\n*\" In the near future \"*\n\nHem, it might depend on the metric." } ] }, { "id": 404005, "postDate": "2018-10-15T05:02:37.900Z", "content": "It's a bit early for me to say this is unfair since I only started this a month ago, but it really means a lot that someone out there is making efforts creating such an environment for us. Thank you!
", "rawMarkdown": "It's a bit early for me to say this is unfair since I only started this a month ago, but it really means a lot that someone out there is making efforts creating such an environment for us. Thank you!", "votes": 2 }, { "id": 403503, "postDate": "2018-10-13T19:00:22.753Z", "content": "people are still busy improving their LB score :P
", "rawMarkdown": "people are still busy improving their LB score :P", "votes": 2 }, { "id": 402652, "postDate": "2018-10-12T04:08:58.687Z", "content": "I think it's fairly simple for RStudio to ask Google for a different Google Analytics Account Data (probably a different region of the world like India's some company) and with this, the key principles went behind the work of many could be retained while also the demo data could not have any impact on this.
", "rawMarkdown": "I think it's fairly simple for RStudio to ask Google for a different Google Analytics Account Data (probably a different region of the world like India's some company) and with this, the key principles went behind the work of many could be retained while also the demo data could not have any impact on this.", "votes": 2, "replies": [ { "id": 402653, "postDate": "2018-10-12T04:09:54.257Z", "content": "Also, that GA data without customer information is anonymized so shouldn't be a problem with respect to data regulation.
", "rawMarkdown": "Also, that GA data without customer information is anonymized so shouldn't be a problem with respect to data regulation. " } ] }, { "id": 402599, "postDate": "2018-10-12T01:18:04.403Z", "content": "Thank you for updating, Julia ! But \n\nan update will occur in the near future that will not benefit from the demo account’s available set of data\n\nis obviously tough work for host.
hi
", "rawMarkdown": "hi" }, { "id": 403807, "postDate": "2018-10-14T16:18:45.143Z", "content": "Is it possible to manually screen out submissions that have obvious signs of abusing the leaked data and still award winners based on those who kept to the spirit of the competition?
", "rawMarkdown": "Is it possible to manually screen out submissions that have obvious signs of abusing the leaked data and still award winners based on those who kept to the spirit of the competition? ", "replies": [ { "id": 403828, "postDate": "2018-10-14T17:19:03.410Z", "content": "Then it'll be a competition on who can be the most subtle in abusing the leak.
", "rawMarkdown": "Then it'll be a competition on who can be the most subtle in abusing the leak.", "votes": 14 }, { "id": 403830, "postDate": "2018-10-14T17:22:40.287Z", "content": "I wish this competition does not continue that way!
", "rawMarkdown": "I wish this competition does not continue that way!", "votes": 1 }, { "id": 403835, "postDate": "2018-10-14T17:28:15.893Z", "content": "Nick, It will not be possible to detect, since we are only submitting the predictions file.
", "rawMarkdown": "Nick, It will not be possible to detect, since we are only submitting the predictions file.", "votes": 2 }, { "id": 403836, "postDate": "2018-10-14T17:31:05.800Z", "content": "@dmitriyguller do you mean having a model that would predict if a kaggker used the real target data...
\n\nif LB_SCORE == 1.30:\n this is suspicious\nelif LB_SCORE == 1.10:\n Are you testing us \nelif LB_SCORE == 0.55:\n you must be Jérôme, aren't you ?\nelif LB_SCORE == 0 :\n Gosh you found the second google demo account\n:-)
", "rawMarkdown": ":-)" }, { "id": 403918, "postDate": "2018-10-14T22:23:51.403Z", "content": "this is really funny, can't stop laughing :)
", "rawMarkdown": "this is really funny, can't stop laughing :)" }, { "id": 404311, "postDate": "2018-10-15T14:54:54.560Z", "content": "Couldn't stop laughing for a long long time :) Thank you for your code..it made my day.
", "rawMarkdown": "Couldn't stop laughing for a long long time :) Thank you for your code..it made my day." } ] }, { "id": 402784, "postDate": "2018-10-12T10:09:06.517Z", "content": "I have just now joined the competition . i am not able to see the \"thatransactionRevenue\"\" sub coloumn in the totals coloumn. i just now downloaded the train.csv into my local machine.
", "rawMarkdown": "I have just now joined the competition . i am not able to see the \"thatransactionRevenue\"\" sub coloumn in the totals coloumn. i just now downloaded the train.csv into my local machine.\n ", "votes": -4, "replies": [ { "id": 403020, "postDate": "2018-10-12T18:14:22.703Z", "content": "totals.transactionRevenue isn't present in 98.6% of the rows. It means that the customer didn't generate any revenue.
", "rawMarkdown": "totals.transactionRevenue isn't present in 98.6% of the rows. It means that the customer didn't generate any revenue.", "votes": 3 }, { "id": 404402, "postDate": "2018-10-15T16:46:49.693Z", "content": "The Customer didn't generate any revenue or they are not sure whether the customer generated the revenue or not.
\n\n\"The customer didn't generate any revenue\" - this statement might be misleading. Please correct me if I am wrong.
", "rawMarkdown": "The Customer didn't generate any revenue or they are not sure whether the customer generated the revenue or not. \n\n\"The customer didn't generate any revenue\" - this statement might be misleading. Please correct me if I am wrong." }, { "id": 405100, "postDate": "2018-10-16T21:38:08.957Z", "content": "It just means that the customer didn't buy anything. I doesn't mean that they are not sure whether the customer generated the revenue or not. \nYou can set totals.transactionRevenue to zero in the rows where it is missing.
", "rawMarkdown": "It just means that the customer didn't buy anything. I doesn't mean that they are not sure whether the customer generated the revenue or not. \nYou can set totals.transactionRevenue to zero in the rows where it is missing. ", "replies": [ { "id": 2506985, "postDate": "2023-10-31T17:10:35.043Z", "content": "make sense
", "rawMarkdown": "make sense" } ] } ] }, { "id": 3168758, "postDate": "2025-04-02T18:49:57.213Z", "content": "Will it still be held this year
", "rawMarkdown": "Will it still be held this year" }, { "id": 2506984, "postDate": "2023-10-31T17:10:04.067Z", "content": "Hi everyone… I am happy to be part of kaggle, and looking forward for being a competent participant
", "rawMarkdown": "Hi everyone... I am happy to be part of kaggle, and looking forward for being a competent participant" }, { "id": 538220, "postDate": "2019-05-28T10:10:01Z", "content": "hi # Julia\nI have a question about competitions \nIs that possible that u tell me how kaggle pays the prizes to the winners of a competition \ndoes it use credit methods like paypal or visa or something like this?\nthanks for your attention
", "rawMarkdown": "hi # Julia\nI have a question about competitions \nIs that possible that u tell me how kaggle pays the prizes to the winners of a competition \ndoes it use credit methods like paypal or visa or something like this?\nthanks for your attention" }, { "id": 424821, "postDate": "2018-11-20T17:58:23.763Z", "content": "Hey, can the ground truth of test_v1.csv be shared to check our previous approaches?
", "rawMarkdown": "Hey, can the ground truth of test_v1.csv be shared to check our previous approaches?", "replies": [ { "id": 424824, "postDate": "2018-11-20T18:06:31.550Z", "content": "Hi Pragya,
\n\nThe ground truth from the first data release is available in train_v2.csv.
\n\nThanks!
", "rawMarkdown": "Hi Pragya,\n\nThe ground truth from the first data release is available in train_v2.csv.\n\nThanks!" } ] }, { "id": 417157, "postDate": "2018-11-07T21:31:57.497Z", "content": "@Kaggle\nWould you mind extending by extra 1-2 weeks? It's not clear why there has been no additional communication since Oct 15. I'm sure there is good reason for the last minute update, but it will have been 24 days since the last update when the data v2 finally arrives. 4 weeks would have been a better estimate.
\n\n\"datasets will be updated in 2-3 weeks time, but no later than November 9, 2018\"
", "rawMarkdown": "@Kaggle\nWould you mind extending by extra 1-2 weeks? It's not clear why there has been no additional communication since Oct 15. I'm sure there is good reason for the last minute update, but it will have been 24 days since the last update when the data v2 finally arrives. 4 weeks would have been a better estimate. \n\n\"datasets will be updated in 2-3 weeks time, but no later than November 9, 2018\"" }, { "id": 415590, "postDate": "2018-11-05T11:15:18.373Z", "content": "Hi Julia Elliott,why the submission system is wrong?
", "rawMarkdown": "Hi Julia Elliott,why the submission system is wrong?" }, { "id": 412282, "postDate": "2018-10-29T23:00:53.880Z", "content": "Anybody having issue with opening the CSV?
", "rawMarkdown": "Anybody having issue with opening the CSV?" }, { "id": 410844, "postDate": "2018-10-26T18:41:48.347Z", "content": ":)
", "rawMarkdown": ":)" }, { "id": 409258, "postDate": "2018-10-24T02:28:07.793Z", "content": "I am thinking is it still worth to take a look on old dataset? Should I just wait for the new datasets?
", "rawMarkdown": "I am thinking is it still worth to take a look on old dataset? Should I just wait for the new datasets?", "replies": [ { "id": 411039, "postDate": "2018-10-27T09:04:35.497Z", "content": "Just get the rough overview of the old data then wait for the new one!
", "rawMarkdown": "Just get the rough overview of the old data then wait for the new one!" } ] }, { "id": 408242, "postDate": "2018-10-22T14:29:14.147Z", "content": "When new data will be released ?
", "rawMarkdown": "When new data will be released ?", "replies": [ { "id": 408427, "postDate": "2018-10-22T21:28:07.110Z", "content": "In the \"data\" tab they say Nov. 9
", "rawMarkdown": "In the \"data\" tab they say Nov. 9" } ] }, { "id": 407431, "postDate": "2018-10-21T06:31:03.753Z", "content": "I think that there is a fundamental problem with the change of the rules:\nUntil the change, we were supposed to take predictions to the target value based on behaviours such as browser-using and advertising affects.\nBut now, we need to forecast the target value of visitors regardless of the properties of the specific session.\nWhat that means, is that the target value in the test is not derived from the features we have (and will have) in the train data, unless we make the very strong (and wrong) assumption that the features would remain constant for the visitorsID. As a result, the best prediction to my opinion would just be the simple average of the target value for each visitor in the past visits!\nI think that a better solution would be that kagglers would not give their predictions for future visitorsID, but rather give their best models (that would start from the raw data file and produce the final prediction for each row in the test data), and the judges would then run the models on the test data.
\n\nBTW: I apologize if my English is not perfect. I come from the (true) middle-east.
", "rawMarkdown": "I think that there is a fundamental problem with the change of the rules:\nUntil the change, we were supposed to take predictions to the target value based on **behaviours** such as browser-using and advertising affects.\nBut now, we need to forecast the target value of **visitors** regardless of the properties of the specific session.\nWhat that means, is that the target value in the test is not derived from the features we have (and will have) in the train data, unless we make the very strong (and wrong) assumption that the features would remain constant for the visitorsID. As a result, the best prediction to my opinion would just be the simple average of the target value for each visitor in the past visits!\nI think that a better solution would be that kagglers would not give their predictions for future visitorsID, but rather give their best models (that would start from the raw data file and produce the final prediction for each row in the test data), and the judges would then run the models on the test data.\n\nBTW: I apologize if my English is not perfect. I come from the (true) middle-east.", "replies": [ { "id": 407459, "postDate": "2018-10-21T08:21:21.360Z", "content": "Average of target value for each visitor is a bit too simple as there are also buying frequencies and trends for each user and we have demographic groups (and characteristics such as mobile, browser; not sure if those count as demographics ;-)), but all session related stuff becomes useless indeed.
\n\nBTW: it becomes more interesting again if we get the full demographics that Google Analytics stores I think. For instance, I think Google Analytics also stores age estimated by the doubleclick tracking cookie.
", "rawMarkdown": "Average of target value for each visitor is a bit too simple as there are also buying frequencies and trends for each user and we have demographic groups (and characteristics such as mobile, browser; not sure if those count as demographics ;-)), but all session related stuff becomes useless indeed.\n\nBTW: it becomes more interesting again if we get the full demographics that Google Analytics stores I think. For instance, I think Google Analytics also stores age estimated by the doubleclick tracking cookie.", "votes": 1 }, { "id": 407495, "postDate": "2018-10-21T10:15:14.423Z", "content": "Hi @Arye,\nI disagree,\nIndeed it's much harder to predict this target than the previous one.\nBut I think that a classification(!) whether a visitor is going to visit\\buy in the test period, will be a key approach in this solution, and there are quite a few interesting features to explore in that area.
", "rawMarkdown": "Hi @Arye,\nI disagree,\nIndeed it's much harder to predict this target than the previous one.\nBut I think that a classification(!) whether a visitor is going to visit\\buy in the test period, will be a key approach in this solution, and there are quite a few interesting features to explore in that area.", "votes": 1 }, { "id": 415766, "postDate": "2018-11-05T16:54:42.957Z", "content": "I'm looking forward to seeing how a benchmark all 0's prediction will fare.
", "rawMarkdown": "I'm looking forward to seeing how a benchmark all 0's prediction will fare." } ] }, { "id": 406999, "postDate": "2018-10-20T06:26:38.287Z", "content": "Hello Kaggle
", "rawMarkdown": "Hello Kaggle" }, { "id": 406614, "postDate": "2018-10-19T14:28:01.643Z", "content": "Thanks for addressing the issue and update. Finally 'real-world' problem!
", "rawMarkdown": "Thanks for addressing the issue and update. Finally 'real-world' problem! " }, { "id": 405565, "postDate": "2018-10-17T18:03:44.630Z", "content": "Sounds like no new columns were added, only removed, is that correct?
", "rawMarkdown": "Sounds like no new columns were added, only removed, is that correct?" }, { "id": 404759, "postDate": "2018-10-16T10:54:41.400Z", "content": "I was parallely working on two competitions, this one and other in Hackerearth. Both of them ended with data leakage.\nThe Hackerearth test data had 99% data copied from training set , resulting in perfect AUC score of .9999 on leaderboard\nFunny thing is that in spite of bringing it to their notice by few of us, a week ago, they are altogether denying the leak ignoring all the proofs.\nAt least here we have acknowledgement of leak
", "rawMarkdown": "I was parallely working on two competitions, this one and other in Hackerearth. Both of them ended with data leakage.\nThe Hackerearth test data had 99% data copied from training set , resulting in perfect AUC score of .9999 on leaderboard\nFunny thing is that in spite of bringing it to their notice by few of us, a week ago, they are altogether denying the leak ignoring all the proofs.\nAt least here we have acknowledgement of leak" }, { "id": 404725, "postDate": "2018-10-16T09:33:23.330Z", "content": "Mark
", "rawMarkdown": "Mark" }, { "id": 404706, "postDate": "2018-10-16T08:43:43.570Z", "content": "I expect that we will soon receive positive news and a revision of the deadline and rules.
", "rawMarkdown": "I expect that we will soon receive positive news and a revision of the deadline and rules. " }, { "id": 404628, "postDate": "2018-10-16T05:22:23.193Z", "content": "I left the Airbus competition in between (when the test image leak was reported) to join this one hoping not to have same issue here...now that this one also had a leak, I might as well go back to Airbus (that one claims to have fixed the leak) :) Hopefully this competition will provide additional time once the issues is fixed
", "rawMarkdown": "I left the Airbus competition in between (when the test image leak was reported) to join this one hoping not to have same issue here...now that this one also had a leak, I might as well go back to Airbus (that one claims to have fixed the leak) :) Hopefully this competition will provide additional time once the issues is fixed\n" }, { "id": 404535, "postDate": "2018-10-15T23:15:15.160Z", "content": "Nice I got this update early and hoping for a revision soon.
", "rawMarkdown": "Nice I got this update early and hoping for a revision soon." }, { "id": 404532, "postDate": "2018-10-15T23:09:36.973Z", "content": "Thanks for the update. I hope you update the new competition soon
", "rawMarkdown": "Thanks for the update. I hope you update the new competition soon" }, { "id": 404284, "postDate": "2018-10-15T14:17:26.420Z", "content": "I am not sure how any of the test data can be salvaged. You would have to remove visitID, which is the timestamp thus a valuable piece of information. You would also have to remove pageviews and hits, which are the most valuable columns in the data. If someone manages to scrape all the session level data from Analytics, they can likely link it to every session in the test data. \nI am concerned that Kaggle will try to create a new dataset that ends up being leaky anyway
", "rawMarkdown": "I am not sure how any of the test data can be salvaged. You would have to remove visitID, which is the timestamp thus a valuable piece of information. You would also have to remove pageviews and hits, which are the most valuable columns in the data. If someone manages to scrape all the session level data from Analytics, they can likely link it to every session in the test data. \nI am concerned that Kaggle will try to create a new dataset that ends up being leaky anyway", "replies": [ { "id": 404639, "postDate": "2018-10-16T05:38:13.990Z", "content": "The whole may have to be changed. If you take a closer look at the demo account, you can download the daily user revenues at a level that approximates the user revenues at a visitor level. One of the examples is 'geoNetwork.city' which is a high cardinality feature. All the categorical features may have to be removed to prevent them from being used to join the dataset from the demo account.
\n\nPreventing leakage in this dataset is a hard problem so I can understand why there isn't a further response from them after 4 days.
", "rawMarkdown": "The whole may have to be changed. If you take a closer look at the demo account, you can download the daily user revenues at a level that approximates the user revenues at a visitor level. One of the examples is 'geoNetwork.city' which is a high cardinality feature. All the categorical features may have to be removed to prevent them from being used to join the dataset from the demo account. \n\nPreventing leakage in this dataset is a hard problem so I can understand why there isn't a further response from them after 4 days. ", "votes": 1 } ] }, { "id": 403055, "postDate": "2018-10-12T19:59:52.173Z", "content": "Thanks for the update, i hope something is done to keep things going smoothly, i was really in the competition spirit.
", "rawMarkdown": "Thanks for the update, i hope something is done to keep things going smoothly, i was really in the competition spirit." }, { "id": 402743, "postDate": "2018-10-12T08:26:08.740Z", "content": "Maybe we can just change the batch of data? Like using the most latest data which answers are not available.
", "rawMarkdown": "Maybe we can just change the batch of data? Like using the most latest data which answers are not available. " }, { "id": 402547, "postDate": "2018-10-11T21:20:21.647Z", "content": "As I understand correctly - the current dataset will be changed?
", "rawMarkdown": "As I understand correctly - the current dataset will be changed?" }, { "id": 413552, "postDate": "2018-11-01T06:09:03.433Z", "rawMarkdown": "", "isDeleted": true }, { "id": 405194, "postDate": "2018-10-17T03:34:30.333Z", "rawMarkdown": "", "isDeleted": true }, { "id": 405138, "postDate": "2018-10-17T00:03:11.553Z", "rawMarkdown": "", "isDeleted": true }, { "id": 404983, "postDate": "2018-10-16T17:34:28.477Z", "rawMarkdown": "", "isDeleted": true }, { "id": 404929, "postDate": "2018-10-16T15:30:27.730Z", "rawMarkdown": "", "isDeleted": true }, { "id": 404858, "postDate": "2018-10-16T13:19:52.433Z", "rawMarkdown": "", "isDeleted": true }, { "id": 403447, "postDate": "2018-10-13T16:40:00.233Z", "rawMarkdown": "", "isDeleted": true }, { "id": 403021, "postDate": "2018-10-12T18:15:53.630Z", "rawMarkdown": "", "isDeleted": true }, { "id": 402850, "postDate": "2018-10-12T12:58:34.960Z", "rawMarkdown": "", "isDeleted": true }, { "id": 402597, "postDate": "2018-10-12T01:16:15.463Z", "rawMarkdown": "", "isDeleted": true }, { "id": 402542, "postDate": "2018-10-11T21:12:37.040Z", "content": "Thank you Julia.
", "rawMarkdown": "Thank you Julia.", "votes": 1 }, { "id": 415388, "postDate": "2018-11-05T03:03:22.050Z", "content": "Thanks
", "rawMarkdown": "Thanks " }, { "id": 406244, "postDate": "2018-10-18T23:31:28.853Z", "content": "Thank you for the update
", "rawMarkdown": "Thank you for the update" }, { "id": 406234, "postDate": "2018-10-18T22:41:56.453Z", "content": "thanks for the update
", "rawMarkdown": "thanks for the update" }, { "id": 405177, "postDate": "2018-10-17T02:30:00.720Z", "content": "Thanks for the update
", "rawMarkdown": "Thanks for the update" }, { "id": 404481, "postDate": "2018-10-15T19:30:08.247Z", "content": "Thank you !!!
", "rawMarkdown": "Thank you !!!" } ], "comments": [ { "id": 404989, "author_name": "Julia Elliott", "author_url": "", "post_date": "2018-10-16T17:49:22.597000", "content": "Thank you all for your patience as we chart a way forward. We’ve decided to make the following updates to the competition design.
\n\nWith this design update confirmed, you should be able to continue building your models with the existing data provided. Once the updated datasets are provided, we’ll update this thread and you will have at least 3 weeks time to refine your models with the actual data.
\n\nAgain, thank you for your participation and bearing with us through this challenging competition design update. We’ve done our very best to minimize the disruption to modeling you may already have done, while ensuring the integrity of the competition’s results.
", "votes": 40, "replies": [ { "id": 404994, "author_name": "Chinta", "author_url": "", "post_date": "2018-10-16T17:52:41.757000", "content": "Thanks for the update and the redesign of competition to make it practical/realistic.
", "votes": 0, "replies": [] }, { "id": 405001, "author_name": "Shahebaz Mohammad", "author_url": "", "post_date": "2018-10-16T18:01:36.357000", "content": "Finally an update which resolves but doesn't resolve all of major concerns.
\n\n\n\n\nThe public leaderboard will contain a timeframe whose target variable\n is publicly available via the demo account, to be used as a mechanism\n for honest submitters to use for cross validation.
\n
I agree on the honesty part. What about leaderboard probing?
", "votes": 0, "replies": [] }, { "id": 405005, "author_name": "Leonid Gilburd", "author_url": "", "post_date": "2018-10-16T18:12:20.670000", "content": "Great idea! Thank you for addressing the issue, kaggle! Now it's a 'real-world' problem! No more LB probing!
", "votes": 0, "replies": [] }, { "id": 405011, "author_name": "Kain", "author_url": "", "post_date": "2018-10-16T18:17:32.210000", "content": "Thanks... I pity those poor ignorant dishonest competitors ;)!!!!
", "votes": 4, "replies": [] }, { "id": 405014, "author_name": "Scirpus", "author_url": "", "post_date": "2018-10-16T18:22:09.823000", "content": "Thanks for update - I don't think you needed to add \"honest\" as I believe it to be pejorative and unnecessary. I assume all other external data will be allowed?
", "votes": 2, "replies": [] }, { "id": 405017, "author_name": "", "author_url": "", "post_date": "2018-10-16T18:27:26.560000", "content": "", "votes": 0, "replies": [] }, { "id": 405019, "author_name": "MichaelP", "author_url": "", "post_date": "2018-10-16T18:28:40.357000", "content": "Thanks for the update and the objective of allowing us to re-use some existing work. Does this mean the competition will be kernels only and you will run our kernels again after 31st January 2019 even though we submitted them by November 30th? Or, does it mean that we need to make the prediction for the test period by November 30th without having the session information for sessions from December 1st?
", "votes": 25, "replies": [] }, { "id": 405020, "author_name": "Herra Huu", "author_url": "", "post_date": "2018-10-16T18:29:27.250000", "content": "nvm
", "votes": 1, "replies": [] }, { "id": 405023, "author_name": "mezoganet", "author_url": "", "post_date": "2018-10-16T18:36:02.333000", "content": "@Julia Elliott
\n\nThe Google Merchandise Store Demo Account effectively becomes freely available to use public data, but will no longer provide any additional benefit to final standings.
\n\nIt would be much simpler that Kaggle provides the Merchandise Store Demo instead of making us losing time to find how to get this file. It seems that except Jerome Vallet no one did succeed to find it.
\n\nAs a matter of proof, the second rank today is quite the double of Jerome's score.
\n\nMy advise is : \nNo additional Data OR Available additional Data for everyone.
\n\nThanks a lot for your post !
", "votes": 32, "replies": [] }, { "id": 405035, "author_name": "Allen", "author_url": "", "post_date": "2018-10-16T19:05:06.207000", "content": "Will we get visit information for the testing set, or will we only get the IDs?
", "votes": 1, "replies": [] }, { "id": 405038, "author_name": "nefedov_m", "author_url": "", "post_date": "2018-10-16T19:07:14.223000", "content": "We should probably expect huge shakeup on Feb 8
", "votes": 3, "replies": [] }, { "id": 405046, "author_name": "ChrisCC", "author_url": "", "post_date": "2018-10-16T19:18:01.190000", "content": "Hi Julia, thank you for the updates. Much appreciated! Just a quick question:
\n\nWith the period of the actual private dataset being changed to December 1, 2018 through January 31, 2019, how can Kaggle publish test dataset that is supposed to contain session information such as page views or hits that can only be available in the future?
\n\nThanks,
", "votes": 20, "replies": [] }, { "id": 405049, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2018-10-16T19:26:05.360000", "content": "Please answer @datahobbit and @chriscc questions. This is the most important thing to me, because I don't think we will have information like pageviews, visits, etc. for the future test data. Therefore our validation strategy should change to suppose that we do not know the validation information at all, even if there was a page visit!
", "votes": 14, "replies": [] }, { "id": 405051, "author_name": "Alexander Firsov", "author_url": "", "post_date": "2018-10-16T19:29:15.957000", "content": "Julia, thanks for update. \nAs I understand we will use all data currently available, but because we have no future info, we will blindly predict log(1 + sum(revenue)) for each known fullVisitorId. Am I correct?
", "votes": 0, "replies": [] }, { "id": 405068, "author_name": "Anirudh Prakash", "author_url": "", "post_date": "2018-10-16T20:08:55.730000", "content": "Please answer @mezoganet query. I guess most of us have the same problem.
\n\nI spent hours working around with the Demo Account finding the right dataset but couldn't.
", "votes": 5, "replies": [] }, { "id": 405070, "author_name": "MrOoijer", "author_url": "", "post_date": "2018-10-16T20:17:03.983000", "content": "@Julia - your description leads me to the following remarks.
\n\na. The remaining question is how to combine the competitors predcitions fronm before Nov 30 2018, 23:99 with the data from after Jan 31, 2019, 23:99 . Either we provide the models and Kaggle runs them or Kaggle provides the data and we run them. Or do we just guess before that data even exists? Then this is becoming a lottery.
\n\nb. Your proposal to only score the customers that also were found in the period up to Oct 15 2018 in the data. This profoundly changes the competition goal, because at this moment we did not filter out the non recurring visitors and they were the major part of the test set. The second big diference is that we are now not aiming at 11 months of test data, but only at 2 months that include the Xmas period, which is an outlier period in the current train set. Will we have one or two Xmax periods in the training set? Without much more clarity here we cannot even start thinking about a good modelling strategy...
\n\nc. You use the term \"test set\" in two different way. In your third bullit it is data from Dec 1, 2018 to Jan 31, 2019/ In your fifth bullit it is a a proportions of the data, with unknown start date (yet) but ending Oct 15, 2018. Very inaccurate and confusing.
", "votes": 28, "replies": [] }, { "id": 405073, "author_name": "Kain", "author_url": "", "post_date": "2018-10-16T20:27:15.110000", "content": "And the plot thickens.... ;)
", "votes": 1, "replies": [] }, { "id": 405084, "author_name": "akshay", "author_url": "", "post_date": "2018-10-16T20:55:52.663000", "content": "This is getting more confusing. The objective of this competition seems to be like:\n1) Identify if there is a data leakage\n2) If yes, can you provide the best ways to deal with it? p.s. Only word files accepted.
", "votes": 7, "replies": [] }, { "id": 405091, "author_name": "mezoganet", "author_url": "", "post_date": "2018-10-16T21:08:06.940000", "content": "@MrOoijer
\n\nI have a simple answer to all your complex questions !
\n\nRead the famous Olivier's (@ogrellier) script : I_have_seen_the_future
\n\n:-)
", "votes": 0, "replies": [] }, { "id": 405096, "author_name": "Aguiar", "author_url": "", "post_date": "2018-10-16T21:14:20.480000", "content": "@juliaelliott thanks for the update.
\n\nI also agree with @mrooijer that this problem is becoming unpredictable. I appreciate that this is a more real scenario, but now we have to guess if the customer will comeback in the next two months, if he is going to buy anything and how much he will spend only with previous data..
", "votes": 1, "replies": [] }, { "id": 405108, "author_name": "Phil Culliton", "author_url": "", "post_date": "2018-10-16T21:56:33.170000", "content": "@datahobbit - good question! It'll be the latter. You'll submit predictions for the users in the test set for that time period (December 1st 2018 to January 31st 2019) by November 30th.
", "votes": 2, "replies": [] }, { "id": 405111, "author_name": "Phil Culliton", "author_url": "", "post_date": "2018-10-16T21:58:44.213000", "content": "@pfichou - sure! There's a link to the demo account on the Data page. As soon as you sign into the demo account you should be presented with a dashboard allowing you to access Google Merchandise Store data.
", "votes": 0, "replies": [] }, { "id": 405113, "author_name": "Phil Culliton", "author_url": "", "post_date": "2018-10-16T21:59:42.703000", "content": "Hey @alexfir - that's correct. You'll need to predict the competition target for unseen data.
", "votes": 1, "replies": [] }, { "id": 405114, "author_name": "Phil Culliton", "author_url": "", "post_date": "2018-10-16T22:01:24.373000", "content": "@anirudhprakash - hmm, sorry about that. When I signed into the demo account I was immediately at the dashboard for the Google Merchandise Store data. What are you seeing when you log in?
", "votes": 0, "replies": [] }, { "id": 405115, "author_name": "Phil Culliton", "author_url": "", "post_date": "2018-10-16T22:03:39.173000", "content": "Hey @chriscc - the competition becomes a problem where you're trying to predict the target for unseen data. So you won't have access to the data for December 1st 2018 to January 31st 2019.
", "votes": 1, "replies": [] }, { "id": 405120, "author_name": "S D", "author_url": "", "post_date": "2018-10-16T22:21:12.500000", "content": "Seems like the Kaggle team figured out a creative way to make this competition possible. I didn't think it was possible, but you've come up with a solution!
\n\nA few comments:
\n\n1) The public leaderboard seems like it will be just for fun, with a huge shake up expected in February! Hopefully most people play along.
\n\n2) The models we have built so far are pretty much useless are they rely on future data such as pageviews/# of visits during the test period.
\n\n3) Will we have to make a different submission for the public LB test period and for the real test period?
\n\n4) People who have extracted data from October 15 to Nov 30 will have an advantage. Hopefully a good samaritan makes it available. We might all have to update our models a few minutes before the deadline!
\n\n5) The new problem seems now to be more realistic, and makes more sense from a business point of view. It will also be much harder to make good predictions!
", "votes": 15, "replies": [] }, { "id": 405124, "author_name": "Herra Huu", "author_url": "", "post_date": "2018-10-16T22:28:35.893000", "content": "How are users who don't visit during the test period handled in the scoring process? Completely ignored or is the target set to zero?
", "votes": 9, "replies": [] }, { "id": 405129, "author_name": "Tak", "author_url": "", "post_date": "2018-10-16T22:52:44.447000", "content": "\n\n\nYou will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.
\n
As Aguiar said, will we predict in this competition the users who will be return and buy something without the detail data in the future, include geo-information, user behavior and traffic source?
", "votes": 0, "replies": [] }, { "id": 405153, "author_name": "MrOoijer", "author_url": "", "post_date": "2018-10-17T01:04:29.047000", "content": "so thats a lottery?
\n\nEdit this was a reply to:
\n\ndatahobbit - good question! It'll be the latter. You'll submit predictions for the users in the test set for that time period (December 1st 2018 to January 31st 2019) by November 30th.
", "votes": 5, "replies": [] }, { "id": 405171, "author_name": "Manimaran Subramanian", "author_url": "", "post_date": "2018-10-17T02:03:47.890000", "content": "Thank you Julia for the update...
\n\nAs I understand correctly from your update and subsequent discussions, we are asked to predict visitor's revenue who visited during test period (all fullVisitorIds in test_v2.csv) i.e., till 15th of Oct'2018. I also guess our submitted model will not be re-run during assessment period. I have following question...
\n\nCustomer buying cycle is a sequence of events i.e., awareness -> Research -> Comparison -> Purchase -> Retention. However, model should predict visitor's 'Purchase' or 'Retention' events without looking customer's behavior?
", "votes": 1, "replies": [] }, { "id": 405220, "author_name": "SimonChen", "author_url": "", "post_date": "2018-10-17T04:33:17.633000", "content": "Based on the data description (from Data Page)
\n\n\n\n\ntrain_v2.csv - the updated training set - contains user transactions from August 1st 2016 to April 30th 2018.
\n \ntest_v2.csv - the updated test set - contains user transactions from May 1st 2018 to October 15th 2018.
\n \nsample_submission_v2.csv - a updated sample submission file in the correct format. Contains all fullVisitorIds in test_v2.csv. Your submission's PredictedLogRevenue column should make forward-looking predictions for each of these fullVisitorIds for the timeframe of December 1st 2018 to January 31st 2019. Review \"What am I predicting?\" above for details.
\n
I have following questions:
\n\nWill test_v2.csv contain \"transactionRevenue\" column ? Since we will use data from 2016-08-01 to 2018-10-15, to predict \"transactionRevenue\" from 2018-12-01 to 2019-01-31.
\n\nWe submit predictions for 2018-12-01 to 2019-01-31 ( private LB ). So how is public LB evaluated ? Do we need to submit two predictions ? One for public LB ( test set period, from 2018-05-01 to 2018-10-15 ), one for private LB ( predictions between 2018-12-01 and 2019-1-31) ?
What about data between 2018-10-16 and 2018-11-30 ? We can't benefit from GA demo account data in this period ?
Thanks for your update.
", "votes": 0, "replies": [] }, { "id": 405312, "author_name": "Yiyang", "author_url": "", "post_date": "2018-10-17T08:10:20.167000", "content": "So I guess we can take a rest for 2 to 3 weeks till the data updated?
", "votes": 1, "replies": [] }, { "id": 405330, "author_name": "narsil (jobs-in-data.com)", "author_url": "", "post_date": "2018-10-17T08:53:16.720000", "content": "@Julia and Kaggle team - big congrats and thank You. What you have essentialy done:
\n\nI think this is very well thought-through. Congrats.
", "votes": 3, "replies": [] }, { "id": 405333, "author_name": "Erik Bruin", "author_url": "", "post_date": "2018-10-17T08:57:33.533000", "content": "@SimonChen. It is really confusing indeed. After a second look it seems like:
\n\ntransactionRevenue for test_v2_LB can probably be mined on Google Analytics. If so, the public LB will be pretty meaningless (but does not count for anything anyway).
\n\nBTW: I don't see that X is May 1st, 2018 on the Data Page at the moment. Where did you find that date?
", "votes": 0, "replies": [] }, { "id": 405352, "author_name": "iliatimofeev", "author_url": "", "post_date": "2018-10-17T10:03:01.990000", "content": "Demo account provides historical data since 1 November 2015. Could you provide supplemental training set for this period? \nAs new task is forecasting Christmas sales volume some additional historical information will be helpful. That will allow to focus on forecasting model rather than scrubbing and cracking demo account.
", "votes": 5, "replies": [] }, { "id": 405429, "author_name": "SimonChen", "author_url": "", "post_date": "2018-10-17T13:20:03.507000", "content": "@Erik Bruin, on Data page, you can see the following description:
\n\n\n\n\ntrain_v2.csv - the updated training set - contains user transactions from August 1st 2016 to April 30th 2018.
\n \ntest_v2.csv - the updated test set - contains user transactions from May 1st 2018 to October 15th 2018.
\n
That means X is May 1st
", "votes": 0, "replies": [] }, { "id": 405614, "author_name": "Rituparna", "author_url": "", "post_date": "2018-10-17T20:11:45.567000", "content": "Thanks for the update @juliaelliott & Kaggle team - Would the validity timeline for the Google Cloud Platform coupons also be extended till the new Competition end date, since currently they are only valid till Nov 19, 2018?
", "votes": 1, "replies": [] }, { "id": 405618, "author_name": "Konrad Banachewicz", "author_url": "", "post_date": "2018-10-17T20:23:55.877000", "content": "@Rituparna: I think it was mentioned elsewhere that it won't be the case, unfortunately.
", "votes": 0, "replies": [] }, { "id": 405622, "author_name": "Rituparna", "author_url": "", "post_date": "2018-10-17T20:28:38.273000", "content": "Thanks for letting me know, Konrad!
", "votes": 0, "replies": [] }, { "id": 405629, "author_name": "olivier", "author_url": "", "post_date": "2018-10-17T20:35:03.203000", "content": "If I understand correctly the new goal is to predict what revenues for certain visitors will be in 9 months time (Train stops in April and predictions have to go through Jan 2019) without knowing sessions ' details from October 15th.
\n\nPeople smart/greedy enough will still be able to grab data from Google Demo Account from May 2018 to Nov 2018...
\n\nAll competitors won't have the same data ?
", "votes": 7, "replies": [] }, { "id": 405635, "author_name": "mezoganet", "author_url": "", "post_date": "2018-10-17T20:39:40.893000", "content": "\n\nGood point !
", "votes": 0, "replies": [] }, { "id": 405911, "author_name": "Misha Lisovyi", "author_url": "", "post_date": "2018-10-18T10:16:31.870000", "content": "@ogrellier indeed, this and also the missing data from Nov 2015 to Jul 2016 goes against the declared intention that
\n\n\n", "votes": 2, "replies": [] }, { "id": 406006, "author_name": "sammy ongaya", "author_url": "", "post_date": "2018-10-18T13:38:00.333000", "content": "The Google Merchandise Store Demo Account effectively becomes freely available to use public data, but will no longer provide any additional benefit to final standings.
\n
Thank you for addressing this issue.
", "votes": 0, "replies": [] }, { "id": 406105, "author_name": "Saravji", "author_url": "", "post_date": "2018-10-18T16:36:52.360000", "content": "Julia & Kaggle team,
\n\nHi @saravji!
\n\n1) I'm glad that people seem to favor the new design. It took some work to figure out what we should do!
\n\n2) Option a is correct.
3) I intend to include the targets in the test set to level the playing field, yep. I consider the public LB to be relatively meaningless already (dependent only on whether people could access the correct parts of the demo account) and we want people to have as much information as possible to solve the new problem formulation.
\n\n4) We did our best to maximize the ability to reuse work - I know that we're still falling short, and I am truly sorry about that. I hope that the new level of realism and learning opportunities help to make up for it.
", "votes": 4, "replies": [] }, { "id": 406272, "author_name": "Riya Gupta", "author_url": "", "post_date": "2018-10-19T00:47:13.127000", "content": "Thanks for the update!
", "votes": 0, "replies": [] }, { "id": 406321, "author_name": "ChrisCC", "author_url": "", "post_date": "2018-10-19T03:43:02.020000", "content": "Hi Phil, thanks for the clarifications which are helpful. Just one more question, the updated data description says:
\n\n\n\n\n** Removed Data Fields**
\n \nSome fields were censored to remove target leakage. The major censored fields are listed below.
\n \n1) hits - This row and nested fields are populated for any and all types of hits. Provides a record of all page visits.
\n \n2) customDimensions - This section contains any user-level or session-level custom dimensions that are set for a session. This is a repeated field and has an entry for each dimension that is set.
\n \n3) totals - Multiple sub-columns were removed from the totals field.
\n
It sounds like hits will be removed but I'm not sure if other attributes under totals, e.g. page views will be removed? Statement 1) and 3) seem inconsistent on page views which is under totals.....
", "votes": 1, "replies": [] }, { "id": 407232, "author_name": "MichaelP", "author_url": "", "post_date": "2018-10-20T17:09:20.763000", "content": "@philculliton - thanks for your response indicating that we will make predictions by Nov 30th for visitor ID's in the test set for the period Dec 1st 2018 to Jan 31st 2019. I understand that this will be based only on visitors who were existing in the period August 1, 2016 to October 15, 2018 (see below quote from Julia's update). Apologies if this has already been clarified and I missed it, but it seems to me that we don't know whether those visitors will visit again in the period Dec 1st 2018 to Jan 31st 2019. Are you saying that the scoring of the private leaderboard will treat all those that don't visit again during that period as having a zero revenue? Or, are you saying that those that don't visit again from Dec 1st 2018 to Jan 31st 2019 will be excluded from the private leaderboard scoring?
\n\n\n", "votes": 7, "replies": [] }, { "id": 407747, "author_name": "Manoj", "author_url": "", "post_date": "2018-10-21T17:47:45.150000", "content": "The test set will be future-looking, composed of all transactions from December 1, 2018 through January 31, 2019. You will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.
\n
it would have been better if the update was separate thread, most people might not even know that there is an update.
", "votes": 3, "replies": [] }, { "id": 408584, "author_name": "Harsh", "author_url": "", "post_date": "2018-10-23T05:46:23.547000", "content": "Thank you very much for update. But what will be meaning to make Google Merchandise Store Demo Account Data available if it will no longer provide any additional benefit to final standings
", "votes": 1, "replies": [] }, { "id": 408601, "author_name": "Ilya Khristoforov", "author_url": "", "post_date": "2018-10-23T06:15:36.247000", "content": "It will be better to make a separate thread when the data and competition description are updated. It is hard to find the updates here...
", "votes": 2, "replies": [] }, { "id": 410417, "author_name": "comsaint", "author_url": "", "post_date": "2018-10-26T02:59:27.090000", "content": "@Julia, I see some issues with the new setting, which I detailed in this post. Do you mind to take a look please?
", "votes": 1, "replies": [] }, { "id": 411354, "author_name": "Dimosthenis Karaflos", "author_url": "", "post_date": "2018-10-27T22:47:30.247000", "content": "First of all, gj at solving the leakage effectivelly. \nAs a matter of fact, I do not understand the reason of a LB in this case. I believe that the test dataset should be included on the training dataset and then make blind predictions for the dates >1/12/2018 for the IDs included on the submission file
", "votes": 2, "replies": [] } ] }, { "id": 404252, "author_name": "Fatih Öztürk", "author_url": "", "post_date": "2018-10-15T13:27:55.923000", "content": "Hi Julia. I know it was because of the weekend, but it's been 4 days and there is still no any official words coming from Kaggle. I expect you to understand how waiting without any information can be annoying for us. Before tomorrow, it would be great to hear some updates like 'we're still working on' or 'we are discussing this and this...' etc.
", "votes": 29, "replies": [ { "id": 404715, "author_name": "Erik Bruin", "author_url": "", "post_date": "2018-10-16T09:02:45.710000", "content": "Indeed, and I have seen this before from Kaggle team. What's difficult about saying something like \"we will have an update for you on Wednesday or so\". Staying totally silent for 5 days is not good.
", "votes": 5, "replies": [] } ] }, { "id": 402528, "author_name": "Leonid Gilburd", "author_url": "", "post_date": "2018-10-11T20:39:48.943000", "content": "Weekend off!!! Hooray!!
", "votes": 21, "replies": [] }, { "id": 402730, "author_name": "Scirpus", "author_url": "", "post_date": "2018-10-12T08:05:44.493000", "content": "One last thing - hopefully when you restart the competition this won't mean Olivier and others lose there kernel medals through no fault of their own. I know Olivier is one gold away from kernels grandmaster status at the minute and it would be tragic if his hard work gets punished.
", "votes": 22, "replies": [ { "id": 402735, "author_name": "Erik Bruin", "author_url": "", "post_date": "2018-10-12T08:12:50.090000", "content": "Congrats to Olivier anyway.......4 Gold medals on the competition is incredible! I \"only\" have one, but want to keep that one of course ;-). Although I cannot imagine this happening (we got the votes!), maybe Kaggle could decouple the kernels from the competition if things change drastically (convert them to dataset kernels).
\n\nBTW: I am also going on strike. See no point in continuing if optimization might be useless anyway after changes to the competition.......
", "votes": 5, "replies": [] }, { "id": 402744, "author_name": "Indranil Bhattacharya", "author_url": "", "post_date": "2018-10-12T08:26:35.320000", "content": "@scirpus, I completely agree with you. However, I feel one can also argue in the same way that he/she worked really hard (either locally and did not submit or submit and was in the top) and his/her hard work was not paid off.
\n\nI guess the optimization the kaggle should plan is to respect hard work of one month of many kagglers and also bring back the competition spirit. So, can there be any middle ground like - kernel medels, kernel points and let's say 0.5X competition points after freezing the lb as of day before yesterday and again relauch the comp with the same spirit as what they are maybe planning to do now.
", "votes": 3, "replies": [] }, { "id": 402753, "author_name": "juliencs", "author_url": "", "post_date": "2018-10-12T08:47:27.363000", "content": "Any method that freezes the leaderboard is not acceptable. You cannot punish people for following the announced timeline and working locally without frantically submitting everyday.
", "votes": 10, "replies": [] }, { "id": 402832, "author_name": "Allen", "author_url": "", "post_date": "2018-10-12T12:29:02.880000", "content": "Maybe they can just close this competition and then launch a new but similar competition. That might make it easy Olivier and others to retain their kernel medals.
", "votes": 1, "replies": [] } ] }, { "id": 402699, "author_name": "Scirpus", "author_url": "", "post_date": "2018-10-12T06:45:22.433000", "content": "I do hope you are at least think of giving Jerome a small token of your appreciation. This will encourage leaks to be found sooner rather than later.
\n\nAlso your nice banner image above clearly shows something that looks like the google store demo ;)
", "votes": 20, "replies": [ { "id": 402710, "author_name": "Jerome Vallet", "author_url": "", "post_date": "2018-10-12T07:23:47.920000", "content": "@Scirpus thank you for all your comments!\nJulia Elliott and Phil Culliton thanked me personally and the Kaggle team has just given me a Kaggle Swag :)\nThank you all for your kindness!
", "votes": 40, "replies": [] }, { "id": 402713, "author_name": "Scirpus", "author_url": "", "post_date": "2018-10-12T07:26:02.463000", "content": "Brilliant News - you thoroughly deserve it!\nWell done to Kaggle too!
", "votes": 2, "replies": [] } ] }, { "id": 402533, "author_name": "olivier", "author_url": "", "post_date": "2018-10-11T20:45:01.390000", "content": "Julia, thanks for the quick update. It's good to know Kaggle is aware and trying to tackle the issue.
\n\nI really hope you'll be able to modify the challenge in a way ground truth data becomes useless but IMHO this would require significant changes in the terms of the competitions and certainly a lot more than a slight change in the metric or/and the target.
\n\nAs far as I'm concerned I don't feel I can go on with my own work on this competition while Kaggle is trying to come up with a contingency plan and act like nothing happened.
", "votes": 16, "replies": [ { "id": 402543, "author_name": "Boris Marjanovic", "author_url": "", "post_date": "2018-10-11T21:16:35.813000", "content": "The solution is simple: don't allow the use of external data. This is a ML competition not a leak hunting competition.
", "votes": 14, "replies": [] }, { "id": 402546, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2018-10-11T21:20:20.027000", "content": "It is funny, because this competition originally started as \"No external data allowed\", but the Kagglers convinced them to change this decision, because we wanted to add information like longitude/latitude, GDP of country, timezone, etc.
\n\nNow, we want to change it back to kernels only/no external data allowed.
", "votes": 6, "replies": [] }, { "id": 402552, "author_name": "Aguiar", "author_url": "", "post_date": "2018-10-11T21:43:38.227000", "content": "I think there is no way to make sure that people are not using external data (except with kernels only and top 3 teams). You can just add an external feature to one model which will enter in a 20-model ensemble and kaggle will never know... so allowing external data was a good move (in my opinion). The problem is that they used a subset of features/examples from public available data, which also include the target variable
", "votes": 5, "replies": [] }, { "id": 402555, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2018-10-11T21:52:59.087000", "content": "@jsaguiar It's just a little confusing to know where the line is drawn. For example, in my model I have labeled English speaking countries based on a Wikipedia page of English speaking countries. I'm not importing any external data, but I am referencing external data this way, without doing a merge between train/test and external data source (e.g. I am doing tr_te[Country=='United States']$english_speaking <- 1; tr_te[Country=='Australia']$english_speaking <- 1; etc.)
So....am I using external data?
\n\nYou can extend this to non-toy examples. I don't want to give away other ones so I can preserve a competitive advantage, but you can probably think of other examples where it gets unclear...
\n\nHow do I know what is safe or not? I emailed Phil about this a couple of days ago but haven't received an answer back, and I don't think they will give me an answer because it's too hard of a question to answer/too subjective.
", "votes": 9, "replies": [] }, { "id": 402571, "author_name": "Aguiar", "author_url": "", "post_date": "2018-10-11T23:23:56.220000", "content": "Exactly, the only way to know if someone used external data is by looking at the code. That's why I think datasets used for competitions should be proprietary (like in Home Credit) or use some method to prevent cheating (two sigma)
", "votes": 1, "replies": [] } ] }, { "id": 404739, "author_name": "Scirpus", "author_url": "", "post_date": "2018-10-16T10:16:30.560000", "content": "Kaggle - you are in danger of losing the goodwill of the competitors\nIn this case No News is definitely not Good News\nWhilst we all realise you are working hard behind the scenes there really is no excuse for zero updates
", "votes": 14, "replies": [ { "id": 404930, "author_name": "akshay", "author_url": "", "post_date": "2018-10-16T15:32:37.710000", "content": "It has become like a TV series, they leave you with a glitch at the end of the show and then you keep on searching/thinking as to what is the next thing that's going to happen. The only difference is that you are aware of the exact dates there as to when it would be aired which doesn't seem to be the case with Competition considering the updates. It's really sad to see that even after seeing such a keen interest from the participants considering the submissions everyone has made they don't find it worthy for a clear answer.
", "votes": 7, "replies": [] }, { "id": 405123, "author_name": "Raja CSP", "author_url": "", "post_date": "2018-10-16T22:27:29.150000", "content": "Kagglers like @akshay are great at comparing real scenarios and observing things!\nKudos!
", "votes": 0, "replies": [] } ] }, { "id": 403066, "author_name": "Indranil Bhattacharya", "author_url": "", "post_date": "2018-10-12T20:11:36.207000", "content": "I don't know why I keep coming to this thread. Maybe this is the 50th time I am visiting this thread today to check if there is something from Kaggle :(
\n\nMaybe, I need some break.
", "votes": 13, "replies": [ { "id": 403097, "author_name": "Aleksandr Kosolapov", "author_url": "", "post_date": "2018-10-12T21:14:02.543000", "content": "", "votes": 11, "replies": [] }, { "id": 403101, "author_name": "Darryl", "author_url": "", "post_date": "2018-10-12T21:32:02.193000", "content": "Also of interest is how long before other kagglers catch up to Jerome
", "votes": 1, "replies": [] } ] }, { "id": 405374, "author_name": "juliencs", "author_url": "", "post_date": "2018-10-17T11:16:33.787000", "content": "My main takeaways :
\n\nAll work until now is not completely wasted, but a large part of it is now irrelevant : study of various CV strategies, most of the feature engineering, study of the impact of each feature (which was heavily linked to the way the problem was framed), etc.
The new public leaderboard will also be largely irrelevant. It covers a very different time period from the new private leaderboard (which covers Christmas aka an outlier period). Also the public leaderboard target will be available on the Demo Account, for this reason you'll be able to simulate the public leaderboard locally without being limited to 5 submissions a day. In effect, the public leaderboard becomes yet another possible local CV.
Predicting only based on past history, without the session-level info for the private leaderboard time period, seems like a quite impossible task, considering once again that this is the Christmas period. Someone mentioned a \"lottery\", I'll wait to see what my CV scores are, but I'm not confident it's gonna be any good.
The Demo Account will be a mandatory data source if you have any hope of competing, for at least 3 reasons :
1- Extract the target for the test set/public leaderboard period, in order to not be limited to 5 submissions a day and/or to use it as part of your train set.
\n\n2- Extract data from November 1 2015 to August 1, 2016 to augment the dataset with a 3rd Christmas period.
\n\n3- Extract data from October 15, 2018 to November 30, 2018 to augment the dataset with 15 possibly precious days of data, the closest ones to the private leaderboard period.
\n\nEdit : Persons posting public kernels with the result of these 3 extractions will get an upvote and infinite good karma coming from me at least :)
", "votes": 12, "replies": [ { "id": 405709, "author_name": "LDMTWO", "author_url": "", "post_date": "2018-10-18T01:04:41.320000", "content": "@timofeev and @juliencs, can you comment on where this 2015 data is? In the User Explorer, I can't view past Mar 9, 2016. Thanks.
", "votes": 0, "replies": [] }, { "id": 405871, "author_name": "iliatimofeev", "author_url": "", "post_date": "2018-10-18T08:56:09.817000", "content": "Home -> custom date range on a first chart\n
1- Extract the target for the test set/public leaderboard period, in order to not be limited to 5 submissions a day and/or to use it as part of your train set.\n\n2- Extract data from November 1 2015 to August 1, 2016 to augment the dataset with a 3rd Christmas period.\n\n3- Extract data from October 15, 2018 to November 30, 2018 to augment the dataset with 15 possibly precious days of data, the closest ones to the private leaderboard period.\nExcellent summary, @juliencs.
\n\nSince this competition got a bit more lottery-like, I propose Kaggle goes full throttle on the idea and allow more final submissions if you pay for them. You know, like in a real lottery, the more coupons you buy the more chances you have. All the money then could be used to hire some more people so that next time all the data for a competion will be available on the site, not to be mined by community. This is just me whining :)
", "votes": 3, "replies": [] }, { "id": 406098, "author_name": "AmirH", "author_url": "", "post_date": "2018-10-18T16:23:18.643000", "content": "@juliencs those with demo account data also have access to additional features that were removed in the competition dataset
", "votes": 0, "replies": [] } ] }, { "id": 403557, "author_name": "Liana Napalkova", "author_url": "", "post_date": "2018-10-13T23:51:17.903000", "content": "Since this is not the first time when a leak occurs, maybe it makes sense to introduce a new type of competition called \"reveal hidden leaks\" :)
", "votes": 11, "replies": [] }, { "id": 406365, "author_name": "Dmitriy Guller", "author_url": "", "post_date": "2018-10-19T06:09:43.330000", "content": "I don't see much point in removing variables due to target leakage now, we're literally predicting the future in the redesigned competition. Excluding them would reward those who go and get them on their own.
", "votes": 9, "replies": [ { "id": 406454, "author_name": "Konrad Banachewicz", "author_url": "", "post_date": "2018-10-19T09:21:07.807000", "content": "Do you mean stuff like pageviews and hits for the public LB period?
", "votes": 0, "replies": [] }, { "id": 406651, "author_name": "Dmitriy Guller", "author_url": "", "post_date": "2018-10-19T15:23:51.173000", "content": "I meant the fields referred to in the \"** Removed Data Fields**\" section. It made sense when this competition was peeking into the future, but not now.
\n\nHow you treat the Public LB is a trickier question; you don't want to make it completely uncorrelated to the Private LB, but on the other hand, the data for the Public LB period would be useful for training the models. If some people can get it anyway, then maybe we should just make it available to everyone and let those who care get a perfect score on the Public LB.
", "votes": 1, "replies": [] }, { "id": 406666, "author_name": "juliencs", "author_url": "", "post_date": "2018-10-19T15:45:58.610000", "content": "I completely agree.
\n\nI'll wait to hear from @juliaelliott or @philculliton about the exact nature of the v2 dataset. It's clear by now that in this new competition, the \"removed data fields\", the \"removed date period\" and the \"target for the whole train+test period\" are pretty much mandatory in order to compete seriously. I'm hoping to spend my time practicing ML, not learning to code bots scraping HTML data. If it turns out the new dataset is not comprehensive and you need to do the demo account scraping by yourself, I'll give up on this competition and look at the other ones available.
\n\nHopefully we can get this information (not even the dataset, just the information) sooner rather than later.
", "votes": 5, "replies": [] }, { "id": 406760, "author_name": "Phil Culliton", "author_url": "", "post_date": "2018-10-19T20:00:43.890000", "content": "Hi @dmitriyguller and @juliencs -
\n\nI agree - I plan to include all of the previously removed data fields in the v2 dataset.
", "votes": 9, "replies": [] }, { "id": 406931, "author_name": "AmirH", "author_url": "", "post_date": "2018-10-20T03:37:57.973000", "content": "That's awesome @philculliton! I strongly support this decision. \nWill you be able to release a small dataset in the expected final structure?\nWe will have only 3 weeks with the new dataset until final deadline, to practically start from scratch with new data, and a new goal, almost none of what was done up to now is relevant, that's a very short time.\nProlonging the competition or releasing part of the data in the final structure would make it easier to come up with a good solution, especially for those of us with less time to do kaggling.
", "votes": 2, "replies": [] }, { "id": 407019, "author_name": "Erik Bruin", "author_url": "", "post_date": "2018-10-20T07:52:59.137000", "content": "I totally agree with Amir. Although I appreciate the new ideas I will pass if this ends up like a mad scramble for three weeks. Why not add another month? That would also take out the Christmas period.
", "votes": 1, "replies": [] } ] }, { "id": 403359, "author_name": "Jonathan Mallia", "author_url": "", "post_date": "2018-10-13T13:05:59.820000", "content": "This competition should be closed as it does not make any more sense to find or exploit already existing leaks simply for the purpose of improving the rank.
", "votes": 9, "replies": [] }, { "id": 404187, "author_name": "Scirpus", "author_url": "", "post_date": "2018-10-15T12:12:35.160000", "content": "I am just hoping they don't rush to such an extent that they swap one leak for another.
", "votes": 7, "replies": [] }, { "id": 402720, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2018-10-12T07:43:28.687000", "content": "Thanks for the update.
\n\n\n\n\nIn the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data.
\n
It is hard to stay motivated and keep working on it with this uncertainty. I feel like our efforts are already wasted.
", "votes": 8, "replies": [ { "id": 402792, "author_name": "narsil (jobs-in-data.com)", "author_url": "", "post_date": "2018-10-12T10:23:42.867000", "content": "Exactly. Until rules are clarified I stop my effort here.
", "votes": 3, "replies": [] }, { "id": 402827, "author_name": "Kain", "author_url": "", "post_date": "2018-10-12T12:09:57.743000", "content": "This sentence about proceeding is meaningless....there is no other way, old data must be switched with properly obfuscated new data.
", "votes": 1, "replies": [] } ] }, { "id": 402630, "author_name": "JohnM", "author_url": "", "post_date": "2018-10-12T03:02:00.380000", "content": "Lots of opinions, so I might as well add to it. I find the whole thing to be amusing and hope we can all keep our sense of humor front and center. Is it true that the 'host team' didn't know of the demo account? The whole purpose of the demo account is to demonstrate the value of Google Analytics! Plus. Erik Bruin came out early with his excellent kernel pointing to the demo account, and I made a direct reference to it in my flow diagram kernel.
\n\nI found the demo account in the second week of the competition, but silly me - the API was locked out and I assumed there was no value to the web interface :( Perhaps I am misunderstanding something - very possible - but it sounds like the people wanting us to analyze Google Analytics data didn't know about the Google Analytics demo. Truth is indeed stranger than fiction!
", "votes": 8, "replies": [ { "id": 402659, "author_name": "echo55555", "author_url": "", "post_date": "2018-10-12T04:33:11.337000", "content": "A lot strange thing could happen in large organization. I have worked with some clients, who did not know how to access a lot of specific information in their own database and need help.
\n\nSome participants guess this competition have other purposes. If the guess was true, Kaggle is taking advantage of us and destructing its own credibility for other company. In my opinion, it is unlikely.
", "votes": 1, "replies": [] }, { "id": 402680, "author_name": "Darryl", "author_url": "", "post_date": "2018-10-12T05:40:40.213000", "content": "I think RStudio wants people to use R as opposed to python. It is fair enough that they don't know about the demo data. I think also it is fair enough that kaggle didn't know about the data. Then there is Google with various departments and products cloud, bigquery, analytics, store. I think the relevant managers at Google should take responsibility. Darryl
", "votes": 3, "replies": [] }, { "id": 402687, "author_name": "Erik Bruin", "author_url": "", "post_date": "2018-10-12T06:05:46.067000", "content": "Thanks for mentioning my kernel John! At, let's say version 2 of my kernel, I already became suspicious to be honest......when I replaced the initial screenshots with screenshots that matched the date range of train+test. I thought: could it, would it......? I eventually decided to not look into it. Also did not really have the time to dig into it. Maybe I should have raised my concerns earlier......
", "votes": 5, "replies": [] } ] }, { "id": 404585, "author_name": "Yao Zhongwei", "author_url": "", "post_date": "2018-10-16T02:39:09.713000", "content": "It's been 4 days already. We need some notice and wish to hear from you, julia.
", "votes": 5, "replies": [] }, { "id": 405699, "author_name": "LDMTWO", "author_url": "", "post_date": "2018-10-18T00:28:45.597000", "content": "By any chance, would anyone (@emorej or @juliaelliott) be willing to provide 1) instructions to get this new REQUIRED data source and 2) provide a dump of the data available up to this month (perhaps Kaggle Datasets or an open git repo)? As I understand, there are 3 years worth, about 30k users, and 4M rows (not sure where I read this; 1 row is day-user-visit summary) and we can only download in batches of 5k from the G-Analytics demo? Further it is not being provided directly by Kaggle (because it will keep getting updated). Somewhere in this tool is the full non aggregated data. Perhaps Kaggle/Google can guide us to siphoning off data to actually make this competition fair?
\n\nEdit: Looks like Mar 9, 2016-present (Where is that 2015 data?). 367K users and up near 1K sessions for some users and up to 128 transactions. For someone really wanting to hunt and download this data to win, they will create a bot that spams Google servers until they get the payload to win the competition with more fine grain info. As an example, I picked out info that I find useful to understand the user and the likeliness of a user spending a certain amount. Although we may not get this in the test set, I could build a model that estimates what will be bought and how many as additional features for the test data if I have this [proposed] training data. It's obviously available, but not readily accessible.
\n\nPurchased Google Bluetooth Headphones for $6,248.75\nAction type\nPAYMENT\nTransaction Id\n30755\nTransaction revenue\n$6,248.75\nTransaction tax\n$437.84\nTransaction shipping\n$6.00\nProduct count\n1\nProduct code\nGGOEGEVB070499\nProduct name\nGoogle Bluetooth Headphones\nProduct revenue\n$6,248.75\nProduct quantity\n125
\n\nEdit 2: After further investigation, I see that the whole G suite of tools provides an impressive set of features. I'm a bit lost in the weeds, but I see that there are 525 fields available. Even latitude/longitude and other \"not in demo\" fields are plainly available now. Most appear to be aggregates, but the cool part (if everyone has fair access) are features like \"avg session duration\" and \"avg time on page\". If only I could build models in the Google suite.
", "votes": 6, "replies": [ { "id": 405719, "author_name": "S D", "author_url": "", "post_date": "2018-10-18T02:01:28.657000", "content": "You have a good point in that the extra data (e.g. what the customer bought) could make a difference!
", "votes": 0, "replies": [] } ] }, { "id": 404684, "author_name": "Stanislav Sopov", "author_url": "", "post_date": "2018-10-16T07:51:21.377000", "content": "Julia please, could you give us new updates!\nWe don't know should we work at the current competition or we need to try something else...
", "votes": 6, "replies": [] }, { "id": 404509, "author_name": "Anirudh Prakash", "author_url": "", "post_date": "2018-10-15T20:51:29.873000", "content": "I keep coming back to this thread hoping for an update. Really confused about which solution to work on. It would be great if we get an update soon.
", "votes": 6, "replies": [] }, { "id": 405057, "author_name": "yrtchn", "author_url": "", "post_date": "2018-10-16T19:40:52.197000", "content": "First of all, thanks for keep this competition going. \nOne (pure hypothetical) thought: how much money will someone's need to do some real buys in GStore between (December 1, 2018) through (January 31, 2019) and use this information to win first place in final leaderboad? Do not predict it, do it (well, some part of it)!
", "votes": 3, "replies": [ { "id": 405122, "author_name": "S D", "author_url": "", "post_date": "2018-10-16T22:24:38.883000", "content": "You would need to have had a session ID in the train period. I do, but unfortunately I might have cleared my cookies recently!
", "votes": 1, "replies": [] } ] }, { "id": 404791, "author_name": "Ranjeet Jain", "author_url": "", "post_date": "2018-10-16T11:47:44.560000", "content": "Its been 4 days. Is there any update ?
", "votes": 3, "replies": [] }, { "id": 403401, "author_name": "Albert Cheng", "author_url": "", "post_date": "2018-10-13T14:48:37.993000", "content": "I believe many kagglers are still working hard on their solution without noticing this update...
", "votes": 4, "replies": [] }, { "id": 402531, "author_name": "AmirH", "author_url": "", "post_date": "2018-10-11T20:43:23.587000", "content": "Thanks @Julia.
\n\n\"you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data\"
\n\nThis means that there's no need to put on the \"Leak Hunter\" hat!\nThat's really encouraging, I hate that hat...
", "votes": 4, "replies": [] }, { "id": 404420, "author_name": "Raja CSP", "author_url": "", "post_date": "2018-10-15T17:11:06.120000", "content": "Will you guys extend the deadline because of the data leak issue?
", "votes": 2, "replies": [] }, { "id": 402557, "author_name": "Randy Boyes", "author_url": "", "post_date": "2018-10-11T21:57:47.967000", "content": "I don't know how you would expect people to just keep working on the current problem when we know the problem will change and have no indication of the type of change to expect
", "votes": 3, "replies": [ { "id": 402575, "author_name": "LDMTWO", "author_url": "", "post_date": "2018-10-11T23:47:40.893000", "content": "because plenty of people were far from the honest top submissions and still have much progress to be made. I do agree with the concern and uncertainty though.
", "votes": 2, "replies": [] } ] }, { "id": 418317, "author_name": "Alexander Firsov", "author_url": "", "post_date": "2018-11-09T16:33:28.897000", "content": "Guys, just in case you wait some kind of notification in this thread about new data availability, it is already here.\nDo not know why organizers did not post update here.
", "votes": 1, "replies": [] }, { "id": 405922, "author_name": "Nikita Varganov", "author_url": "", "post_date": "2018-10-18T10:41:56.627000", "content": "Will you reset the LB?
", "votes": 1, "replies": [] }, { "id": 405617, "author_name": "Konrad Banachewicz", "author_url": "", "post_date": "2018-10-17T20:21:58.207000", "content": "I have to say, based on the description in the \"update\" post, that it might get interesting: the 6 weeks gap between the end of training data and the start of real test raises the difficulty level considerably (unless we all refit our models in the last day and Kaggle servers survive literally everyone having to submit at the last moment).
", "votes": 1, "replies": [] }, { "id": 404164, "author_name": "Gaurav Chavan", "author_url": "", "post_date": "2018-10-15T11:07:07.480000", "content": "I had already started working on it. The main gist is how we contribute to thought process. Data Leakage will not contribute much to Data Science.
", "votes": 1, "replies": [] }, { "id": 402712, "author_name": "", "author_url": "", "post_date": "2018-10-12T07:24:59.510000", "content": "", "votes": 1, "replies": [] }, { "id": 402598, "author_name": "", "author_url": "", "post_date": "2018-10-12T01:17:52.920000", "content": "", "votes": 1, "replies": [] }, { "id": 402550, "author_name": "", "author_url": "", "post_date": "2018-10-11T21:33:55.553000", "content": "", "votes": 1, "replies": [] }, { "id": 417661, "author_name": "", "author_url": "", "post_date": "2018-11-08T15:43:40.497000", "content": "", "votes": 2, "replies": [ { "id": 417675, "author_name": "", "author_url": "", "post_date": "2018-11-08T16:06:54.447000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 417194, "author_name": "", "author_url": "", "post_date": "2018-11-07T23:20:48.340000", "content": "", "votes": 2, "replies": [] }, { "id": 415764, "author_name": "", "author_url": "", "post_date": "2018-11-05T16:51:38.640000", "content": "", "votes": 2, "replies": [] }, { "id": 406083, "author_name": "", "author_url": "", "post_date": "2018-10-18T16:06:42.073000", "content": "", "votes": 2, "replies": [] }, { "id": 405217, "author_name": "", "author_url": "", "post_date": "2018-10-17T04:28:14.977000", "content": "", "votes": 2, "replies": [ { "id": 405253, "author_name": "", "author_url": "", "post_date": "2018-10-17T05:50:43.257000", "content": "", "votes": 1, "replies": [] }, { "id": 405394, "author_name": "", "author_url": "", "post_date": "2018-10-17T12:09:43.747000", "content": "", "votes": 1, "replies": [] }, { "id": 406213, "author_name": "", "author_url": "", "post_date": "2018-10-18T21:24:50.967000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 405063, "author_name": "", "author_url": "", "post_date": "2018-10-16T19:49:29.753000", "content": "", "votes": 2, "replies": [] }, { "id": 404780, "author_name": "", "author_url": "", "post_date": "2018-10-16T11:32:14.233000", "content": "", "votes": 2, "replies": [] }, { "id": 404737, "author_name": "", "author_url": "", "post_date": "2018-10-16T10:08:34.567000", "content": "", "votes": 2, "replies": [ { "id": 404796, "author_name": "", "author_url": "", "post_date": "2018-10-16T11:52:41.743000", "content": "", "votes": 1, "replies": [] }, { "id": 404970, "author_name": "", "author_url": "", "post_date": "2018-10-16T17:12:47.353000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 404005, "author_name": "", "author_url": "", "post_date": "2018-10-15T05:02:37.900000", "content": "", "votes": 2, "replies": [] }, { "id": 403503, "author_name": "", "author_url": "", "post_date": "2018-10-13T19:00:22.753000", "content": "", "votes": 2, "replies": [] }, { "id": 402652, "author_name": "", "author_url": "", "post_date": "2018-10-12T04:08:58.687000", "content": "", "votes": 2, "replies": [ { "id": 402653, "author_name": "", "author_url": "", "post_date": "2018-10-12T04:09:54.257000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 402599, "author_name": "", "author_url": "", "post_date": "2018-10-12T01:18:04.403000", "content": "", "votes": 2, "replies": [] }, { "id": 403120, "author_name": "", "author_url": "", "post_date": "2018-10-12T23:17:18.230000", "content": "", "votes": 0, "replies": [] }, { "id": 403807, "author_name": "", "author_url": "", "post_date": "2018-10-14T16:18:45.143000", "content": "", "votes": 0, "replies": [ { "id": 403828, "author_name": "", "author_url": "", "post_date": "2018-10-14T17:19:03.410000", "content": "", "votes": 14, "replies": [] }, { "id": 403830, "author_name": "", "author_url": "", "post_date": "2018-10-14T17:22:40.287000", "content": "", "votes": 1, "replies": [] }, { "id": 403835, "author_name": "", "author_url": "", "post_date": "2018-10-14T17:28:15.893000", "content": "", "votes": 2, "replies": [] }, { "id": 403836, "author_name": "", "author_url": "", "post_date": "2018-10-14T17:31:05.800000", "content": "", "votes": 48, "replies": [] }, { "id": 403838, "author_name": "", "author_url": "", "post_date": "2018-10-14T17:37:56.270000", "content": "", "votes": 0, "replies": [] }, { "id": 403918, "author_name": "", "author_url": "", "post_date": "2018-10-14T22:23:51.403000", "content": "", "votes": 0, "replies": [] }, { "id": 404311, "author_name": "", "author_url": "", "post_date": "2018-10-15T14:54:54.560000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 402784, "author_name": "", "author_url": "", "post_date": "2018-10-12T10:09:06.517000", "content": "", "votes": -4, "replies": [ { "id": 403020, "author_name": "", "author_url": "", "post_date": "2018-10-12T18:14:22.703000", "content": "", "votes": 3, "replies": [] }, { "id": 404402, "author_name": "", "author_url": "", "post_date": "2018-10-15T16:46:49.693000", "content": "", "votes": 0, "replies": [] }, { "id": 405100, "author_name": "", "author_url": "", "post_date": "2018-10-16T21:38:08.957000", "content": "", "votes": 0, "replies": [ { "id": 2506985, "author_name": "", "author_url": "", "post_date": "2023-10-31T17:10:35.043000", "content": "", "votes": 0, "replies": [] } ] } ] }, { "id": 3168758, "author_name": "", "author_url": "", "post_date": "2025-04-02T18:49:57.213000", "content": "", "votes": 0, "replies": [] }, { "id": 2506984, "author_name": "", "author_url": "", "post_date": "2023-10-31T17:10:04.067000", "content": "", "votes": 0, "replies": [] }, { "id": 538220, "author_name": "", "author_url": "", "post_date": "2019-05-28T10:10:01", "content": "", "votes": 0, "replies": [] }, { "id": 424821, "author_name": "", "author_url": "", "post_date": "2018-11-20T17:58:23.763000", "content": "", "votes": 0, "replies": [ { "id": 424824, "author_name": "", "author_url": "", "post_date": "2018-11-20T18:06:31.550000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 417157, "author_name": "", "author_url": "", "post_date": "2018-11-07T21:31:57.497000", "content": "", "votes": 0, "replies": [] }, { "id": 415590, "author_name": "", "author_url": "", "post_date": "2018-11-05T11:15:18.373000", "content": "", "votes": 0, "replies": [] }, { "id": 412282, "author_name": "", "author_url": "", "post_date": "2018-10-29T23:00:53.880000", "content": "", "votes": 0, "replies": [] }, { "id": 410844, "author_name": "", "author_url": "", "post_date": "2018-10-26T18:41:48.347000", "content": "", "votes": 0, "replies": [] }, { "id": 409258, "author_name": "", "author_url": "", "post_date": "2018-10-24T02:28:07.793000", "content": "", "votes": 0, "replies": [ { "id": 411039, "author_name": "", "author_url": "", "post_date": "2018-10-27T09:04:35.497000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 408242, "author_name": "", "author_url": "", "post_date": "2018-10-22T14:29:14.147000", "content": "", "votes": 0, "replies": [ { "id": 408427, "author_name": "", "author_url": "", "post_date": "2018-10-22T21:28:07.110000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 407431, "author_name": "", "author_url": "", "post_date": "2018-10-21T06:31:03.753000", "content": "", "votes": 0, "replies": [ { "id": 407459, "author_name": "", "author_url": "", "post_date": "2018-10-21T08:21:21.360000", "content": "", "votes": 1, "replies": [] }, { "id": 407495, "author_name": "", "author_url": "", "post_date": "2018-10-21T10:15:14.423000", "content": "", "votes": 1, "replies": [] }, { "id": 415766, "author_name": "", "author_url": "", "post_date": "2018-11-05T16:54:42.957000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 406999, "author_name": "", "author_url": "", "post_date": "2018-10-20T06:26:38.287000", "content": "", "votes": 0, "replies": [] }, { "id": 406614, "author_name": "", "author_url": "", "post_date": "2018-10-19T14:28:01.643000", "content": "", "votes": 0, "replies": [] }, { "id": 405565, "author_name": "", "author_url": "", "post_date": "2018-10-17T18:03:44.630000", "content": "", "votes": 0, "replies": [] }, { "id": 404759, "author_name": "", "author_url": "", "post_date": "2018-10-16T10:54:41.400000", "content": "", "votes": 0, "replies": [] }, { "id": 404725, "author_name": "", "author_url": "", "post_date": "2018-10-16T09:33:23.330000", "content": "", "votes": 0, "replies": [] }, { "id": 404706, "author_name": "", "author_url": "", "post_date": "2018-10-16T08:43:43.570000", "content": "", "votes": 0, "replies": [] }, { "id": 404628, "author_name": "", "author_url": "", "post_date": "2018-10-16T05:22:23.193000", "content": "", "votes": 0, "replies": [] }, { "id": 404535, "author_name": "", "author_url": "", "post_date": "2018-10-15T23:15:15.160000", "content": "", "votes": 0, "replies": [] }, { "id": 404532, "author_name": "", "author_url": "", "post_date": "2018-10-15T23:09:36.973000", "content": "", "votes": 0, "replies": [] }, { "id": 404284, "author_name": "", "author_url": "", "post_date": "2018-10-15T14:17:26.420000", "content": "", "votes": 0, "replies": [ { "id": 404639, "author_name": "", "author_url": "", "post_date": "2018-10-16T05:38:13.990000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 403055, "author_name": "", "author_url": "", "post_date": "2018-10-12T19:59:52.173000", "content": "", "votes": 0, "replies": [] }, { "id": 402743, "author_name": "", "author_url": "", "post_date": "2018-10-12T08:26:08.740000", "content": "", "votes": 0, "replies": [] }, { "id": 402547, "author_name": "", "author_url": "", "post_date": "2018-10-11T21:20:21.647000", "content": "", "votes": 0, "replies": [] }, { "id": 413552, "author_name": "", "author_url": "", "post_date": "2018-11-01T06:09:03.433000", "content": "", "votes": 0, "replies": [] }, { "id": 405194, "author_name": "", "author_url": "", "post_date": "2018-10-17T03:34:30.333000", "content": "", "votes": 0, "replies": [] }, { "id": 405138, "author_name": "", "author_url": "", "post_date": "2018-10-17T00:03:11.553000", "content": "", "votes": 0, "replies": [] }, { "id": 404983, "author_name": "", "author_url": "", "post_date": "2018-10-16T17:34:28.477000", "content": "", "votes": 0, "replies": [] }, { "id": 404929, "author_name": "", "author_url": "", "post_date": "2018-10-16T15:30:27.730000", "content": "", "votes": 0, "replies": [] }, { "id": 404858, "author_name": "", "author_url": "", "post_date": "2018-10-16T13:19:52.433000", "content": "", "votes": 0, "replies": [] }, { "id": 403447, "author_name": "", "author_url": "", "post_date": "2018-10-13T16:40:00.233000", "content": "", "votes": 0, "replies": [] }, { "id": 403021, "author_name": "", "author_url": "", "post_date": "2018-10-12T18:15:53.630000", "content": "", "votes": 0, "replies": [] }, { "id": 402850, "author_name": "", "author_url": "", "post_date": "2018-10-12T12:58:34.960000", "content": "", "votes": 0, "replies": [] }, { "id": 402597, "author_name": "", "author_url": "", "post_date": "2018-10-12T01:16:15.463000", "content": "", "votes": 0, "replies": [] }, { "id": 402542, "author_name": "", "author_url": "", "post_date": "2018-10-11T21:12:37.040000", "content": "", "votes": 1, "replies": [] }, { "id": 415388, "author_name": "", "author_url": "", "post_date": "2018-11-05T03:03:22.050000", "content": "", "votes": 0, "replies": [] }, { "id": 406244, "author_name": "", "author_url": "", "post_date": "2018-10-18T23:31:28.853000", "content": "", "votes": 0, "replies": [] }, { "id": 406234, "author_name": "", "author_url": "", "post_date": "2018-10-18T22:41:56.453000", "content": "", "votes": 0, "replies": [] }, { "id": 405177, "author_name": "", "author_url": "", "post_date": "2018-10-17T02:30:00.720000", "content": "", "votes": 0, "replies": [] }, { "id": 404481, "author_name": "", "author_url": "", "post_date": "2018-10-15T19:30:08.247000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "402520": "**UPDATE: See [update post in thread][1].**\n\nWe’ve been made aware that the [Google Merchandise Store Demo Account][2] serves as an external source for mining the target variable. Despite our best efforts to have eyes on every corner of potential access to a competition's ground truth, this account’s existence was regrettably unknown to both Kaggle and the host team. And we’re deeply sorry for the impact it’s had on the past month of hard work you’ve put in. Thank you to those (including [Jerome][3]) who [alerted us][4] to the compromised access to the test set’s solution. \n\nWe are currently investigating a revision to the original problem which will enable us to keep the competition running and ideally minimize the disruption to your existing work. In the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data. There may be a slight revision to the measurement of the target variable, but we are doing our best to keep in tact the directional intent of the original problem. \n\nPlease trust that coming to a decision on this and updating you is a top priority for us, and the time we are taking is with the hope that this will be the very final change required for this competition. Further updates with the go forward decision will be posted to this thread.\n\n\n [1]: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/68353#404989\n [2]: https://support.google.com/analytics/answer/6367342#access\n [3]: https://www.kaggle.com/emorej\n [4]: https://www.kaggle.com/c/ga-customer-revenue-prediction/discussion/68235#401880", "404989": "Thank you all for your patience as we chart a way forward. We’ve decided to make the following updates to the competition design.\n\n- The competition end date has been extended until November 30, 2018 to compensate for the time required to revise the dataset and design.\n- We will release an updated dataset that extends the timeframe of available data from August 1, 2016 through October 15, 2018. This will be further split between a training set and public leaderboard test set. The public leaderboard will contain a timeframe whose target variable is publicly available via the demo account, to be used as a mechanism for honest submitters to use for cross validation.\n- The test set will be future-looking, composed of all transactions from December 1, 2018 through January 31, 2019. You will be asked to predict the same calculation of the natural log of the sum of all transactions for each existing fullVisitorId, but only for the existing visitors from August 1, 2016 to October 15, 2018.\n- The [Google Merchandise Store Demo Account][1] effectively becomes freely available to use public data, but will no longer provide any additional benefit to final standings.\n- The train, test & sample submission datasets will be updated in 2-3 weeks time, but no later than November 9, 2018. While the time gap is not ideal, it is required for us to pull and format the data needed.\n- The competition’s private leaderboard will be scored and refreshed on February 8, 2019, at which point the standings will be final.\n- The prize money and allocations will remain in tact, but its distribution will be delayed until after the final standings are confirmed and winners’ obligations are fulfilled.\n\nWith this design update confirmed, you should be able to continue building your models with the existing data provided. Once the updated datasets are provided, we’ll update this thread and you will have at least 3 weeks time to refine your models with the actual data.\n\nAgain, thank you for your participation and bearing with us through this challenging competition design update. We’ve done our very best to minimize the disruption to modeling you may already have done, while ensuring the integrity of the competition’s results.\n\n\n[1]: https://support.google.com/analytics/answer/6367342#access", "404252": "Hi Julia. I know it was because of the weekend, but it's been 4 days and there is still no any official words coming from Kaggle. I expect you to understand how waiting without any information can be annoying for us. Before tomorrow, it would be great to hear some updates like 'we're still working on' or 'we are discussing this and this...' etc. ", "402528": "Weekend off!!! Hooray!! ", "402730": "One last thing - hopefully when you restart the competition this won't mean Olivier and others lose there kernel medals through no fault of their own. I know Olivier is one gold away from kernels grandmaster status at the minute and it would be tragic if his hard work gets punished.", "402699": "I do hope you are at least think of giving Jerome a small token of your appreciation. This will encourage leaks to be found sooner rather than later.\n\nAlso your nice banner image above clearly shows something that looks like the google store demo ;)", "402533": "Julia, thanks for the quick update. It's good to know Kaggle is aware and trying to tackle the issue.\n\nI really hope you'll be able to modify the challenge in a way ground truth data becomes useless but IMHO this would require significant changes in the terms of the competitions and certainly a lot more than a slight change in the metric or/and the target.\n\nAs far as I'm concerned I don't feel I can go on with my own work on this competition while Kaggle is trying to come up with a contingency plan and act like nothing happened. ", "404739": "Kaggle - you are in danger of losing the goodwill of the competitors\nIn this case No News is definitely not Good News\nWhilst we all realise you are working hard behind the scenes there really is no excuse for zero updates", "403066": "I don't know why I keep coming to this thread. Maybe this is the 50th time I am visiting this thread today to check if there is something from Kaggle :( \n\nMaybe, I need some break. ", "405374": "My main takeaways :\n\n* All work until now is not completely wasted, but a large part of it is now irrelevant : study of various CV strategies, most of the feature engineering, study of the impact of each feature (which was heavily linked to the way the problem was framed), etc.\n\n* The new public leaderboard will also be largely irrelevant. It covers a very different time period from the new private leaderboard (which covers Christmas aka an outlier period). Also the public leaderboard target will be available on the Demo Account, for this reason you'll be able to simulate the public leaderboard locally without being limited to 5 submissions a day. In effect, the public leaderboard becomes yet another possible local CV.\n\n* Predicting only based on past history, without the session-level info for the private leaderboard time period, seems like a quite impossible task, considering once again that this is the Christmas period. Someone mentioned a \"lottery\", I'll wait to see what my CV scores are, but I'm not confident it's gonna be any good.\n\n* The Demo Account will be a mandatory data source if you have any hope of competing, for at least 3 reasons :\n\n1- Extract the target for the test set/public leaderboard period, in order to not be limited to 5 submissions a day and/or to use it as part of your train set.\n\n2- Extract data from November 1 2015 to August 1, 2016 to augment the dataset with a 3rd Christmas period.\n\n3- Extract data from October 15, 2018 to November 30, 2018 to augment the dataset with 15 possibly precious days of data, the closest ones to the private leaderboard period.\n\nEdit : Persons posting public kernels with the result of these 3 extractions will get an upvote and infinite good karma coming from me at least :)", "403557": "Since this is not the first time when a leak occurs, maybe it makes sense to introduce a new type of competition called \"reveal hidden leaks\" :) ", "406365": "I don't see much point in removing variables due to target leakage now, we're literally predicting the future in the redesigned competition. Excluding them would reward those who go and get them on their own.", "403359": "This competition should be closed as it does not make any more sense to find or exploit already existing leaks simply for the purpose of improving the rank.", "404187": "I am just hoping they don't rush to such an extent that they swap one leak for another. ", "402720": "Thanks for the update.\n\n> In the meantime, you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data.\n\nIt is hard to stay motivated and keep working on it with this uncertainty. I feel like our efforts are already wasted.", "402630": "Lots of opinions, so I might as well add to it. I find the whole thing to be amusing and hope we can all keep our sense of humor front and center. Is it true that the 'host team' didn't know of the demo account? The whole purpose of the demo account is to demonstrate the value of Google Analytics! Plus. Erik Bruin came out early with his [excellent kernel](https://www.kaggle.com/erikbruin/google-analytics-eda-lightgbm-screenshots) pointing to the demo account, and I made a direct reference to it in my [flow diagram kernel](https://www.kaggle.com/jpmiller/go-with-the-flow-lb-1-447). \n\nI found the demo account in the second week of the competition, but silly me - the API was locked out and I assumed there was no value to the web interface :( Perhaps I am misunderstanding something - very possible - but it sounds like the people wanting us to analyze Google Analytics data didn't know about the Google Analytics demo. Truth is indeed stranger than fiction!", "404585": "It's been 4 days already. We need some notice and wish to hear from you, julia.", "405699": "By any chance, would anyone (@emorej or @juliaelliott) be willing to provide 1) instructions to get this new REQUIRED data source and 2) provide a dump of the data available up to this month (perhaps Kaggle Datasets or an open git repo)? As I understand, there are 3 years worth, about 30k users, and 4M rows (not sure where I read this; 1 row is day-user-visit summary) and we can only download in batches of 5k from the G-Analytics demo? Further it is not being provided directly by Kaggle (because it will keep getting updated). Somewhere in this tool is the full non aggregated data. Perhaps Kaggle/Google can guide us to siphoning off data to actually make this competition fair?\n\nEdit: Looks like Mar 9, 2016-present (Where is that 2015 data?). 367K users and up near 1K sessions for some users and up to 128 transactions. For someone really wanting to hunt and download this data to win, they will create a bot that spams Google servers until they get the payload to win the competition with more fine grain info. As an example, I picked out info that I find useful to understand the user and the likeliness of a user spending a certain amount. Although we may not get this in the test set, I could build a model that estimates what will be bought and how many as additional features for the test data if I have this [proposed] training data. It's obviously available, but not readily accessible.\n\n\nPurchased Google Bluetooth Headphones for $6,248.75\nAction type\nPAYMENT\nTransaction Id\n30755\nTransaction revenue\n$6,248.75\nTransaction tax\n$437.84\nTransaction shipping\n$6.00\nProduct count\n1\nProduct code\nGGOEGEVB070499\nProduct name\nGoogle Bluetooth Headphones\nProduct revenue\n$6,248.75\nProduct quantity\n125\n\nEdit 2: After further investigation, I see that the whole G suite of tools provides an impressive set of features. I'm a bit lost in the weeds, but I see that there are 525 fields available. Even latitude/longitude and other \"not in demo\" fields are plainly available now. Most appear to be aggregates, but the cool part (if everyone has fair access) are features like \"avg session duration\" and \"avg time on page\". If only I could build models in the Google suite.", "404684": "Julia please, could you give us new updates!\nWe don't know should we work at the current competition or we need to try something else...", "404509": "I keep coming back to this thread hoping for an update. Really confused about which solution to work on. It would be great if we get an update soon.", "405057": "First of all, thanks for keep this competition going. \nOne (pure hypothetical) thought: how much money will someone's need to do some real buys in GStore between (December 1, 2018) through (January 31, 2019) and use this information to win first place in final leaderboad? Do not predict it, do it (well, some part of it)!", "404791": "Its been 4 days. Is there any update ?", "403401": "I believe many kagglers are still working hard on their solution without noticing this update... ", "402531": "Thanks @Julia.\n\n\"you may proceed with your own work with the expectation that an update will occur in the near future that will not benefit from the demo account’s available set of data\"\n\nThis means that there's no need to put on the \"Leak Hunter\" hat!\nThat's really encouraging, I hate that hat...", "404420": "Will you guys extend the deadline because of the data leak issue? ", "402557": "I don't know how you would expect people to just keep working on the current problem when we know the problem will change and have no indication of the type of change to expect", "418317": "Guys, just in case you wait some kind of notification in this thread about new data availability, it is already here.\nDo not know why organizers did not post update here.", "405922": "Will you reset the LB? ", "405617": "I have to say, based on the description in the \"update\" post, that it might get interesting: the 6 weeks gap between the end of training data and the start of real test raises the difficulty level considerably (unless we all refit our models in the last day and Kaggle servers survive literally everyone having to submit at the last moment). ", "404164": "I had already started working on it. The main gist is how we contribute to thought process. Data Leakage will not contribute much to Data Science.", "402712": "Thank you, Julia!\nI hope you will find a good decision, but i think you should public demo account data anyway. I hope it will be ml contest raither than scrapper contes.", "402598": "This is a leak we already know. What I am worried about is whether there will be other similar leaks like [Google Merchandise Store Demo Account][1] ..can get data \n\n\n [1]: https://support.google.com/analytics/answer/6367342#access", "402550": "Hope that time for this competition will add, if dataset changes)", "417661": "I hope we'll hear from Kaggle with an update soon, but I also hope it's not going to be rushed. If it takes longer than expected, I'd prefer waiting a little more than participating in another flawed competition.", "417194": "Will test_v2.csv contain \"transactionRevenue\" column ?", "415764": "5 days remaining, no news about the expected update", "406083": "Are you going to make sure the test data will be two months only so as to mimic (December 1, 2018) through (January 31, 2019)? We don't want people optimizing models that have vastly more data in test than that of the blind test data. As we are trying to predict the sum of revenues this will reduce the probability of competitors vastly over estimating the revenue if they use 1 model to predict individual entries followed by another predicting the revenue sum. (See olivier's script)", "405217": "Dear @juliaelliott,\nthis is my first ever kaggle competition. I worked on it the last two weekends and have learned a lot.\nBut now I returned to the competition page and almost accidentally find out that everything has changed: in the time since I accepted the competition rules, you have changed the rules, you have changed the deadlines AND you have changed the data.\n\nYou have the email addresses of everyone who has entered the competition -- why do you not notify people about the new rules and data??? I expect the majority of the participants out there is not even aware of the new situation!\n\nI do not consider the time I used on this problem wasted, since it contributed to my learning; but regarding the competition it actually is wasted, and I find the way kaggle treats the participants very demotivating. ", "405063": "I recall someone mentioning that not all transactions were reflected in the modeling dataset, only the latest transaction for the session. Is that the case, and if it is, will it be fixed in the new dataset?", "404780": "Hi Julia,\nPlease inform us ASAP because the saying is right: \"A bad decision is better than no decision\". This 'no decision' situation is frustrating all of us.", "404737": " \" In the near future \" usually means A few days or weeks ? I think it's better to tell us \"updating in about a week\" or something like that. Give us a more specific time information, then we'll be more content with other thing.", "404005": "It's a bit early for me to say this is unfair since I only started this a month ago, but it really means a lot that someone out there is making efforts creating such an environment for us. Thank you!", "403503": "people are still busy improving their LB score :P", "402652": "I think it's fairly simple for RStudio to ask Google for a different Google Analytics Account Data (probably a different region of the world like India's some company) and with this, the key principles went behind the work of many could be retained while also the demo data could not have any impact on this.", "402599": "Thank you for updating, Julia ! But \n```\nan update will occur in the near future that will not benefit from the demo account’s available set of data\n```\nis obviously tough work for host.", "403120": "hi", "403807": "Is it possible to manually screen out submissions that have obvious signs of abusing the leaked data and still award winners based on those who kept to the spirit of the competition? ", "402784": "I have just now joined the competition . i am not able to see the \"thatransactionRevenue\"\" sub coloumn in the totals coloumn. i just now downloaded the train.csv into my local machine.\n ", "3168758": "Will it still be held this year", "2506984": "Hi everyone... I am happy to be part of kaggle, and looking forward for being a competent participant", "538220": "hi # Julia\nI have a question about competitions \nIs that possible that u tell me how kaggle pays the prizes to the winners of a competition \ndoes it use credit methods like paypal or visa or something like this?\nthanks for your attention", "424821": "Hey, can the ground truth of test_v1.csv be shared to check our previous approaches?", "417157": "@Kaggle\nWould you mind extending by extra 1-2 weeks? It's not clear why there has been no additional communication since Oct 15. I'm sure there is good reason for the last minute update, but it will have been 24 days since the last update when the data v2 finally arrives. 4 weeks would have been a better estimate. \n\n\"datasets will be updated in 2-3 weeks time, but no later than November 9, 2018\"", "415590": "Hi Julia Elliott,why the submission system is wrong?", "412282": "Anybody having issue with opening the CSV?", "410844": ":)", "409258": "I am thinking is it still worth to take a look on old dataset? Should I just wait for the new datasets?", "408242": "When new data will be released ?", "407431": "I think that there is a fundamental problem with the change of the rules:\nUntil the change, we were supposed to take predictions to the target value based on **behaviours** such as browser-using and advertising affects.\nBut now, we need to forecast the target value of **visitors** regardless of the properties of the specific session.\nWhat that means, is that the target value in the test is not derived from the features we have (and will have) in the train data, unless we make the very strong (and wrong) assumption that the features would remain constant for the visitorsID. As a result, the best prediction to my opinion would just be the simple average of the target value for each visitor in the past visits!\nI think that a better solution would be that kagglers would not give their predictions for future visitorsID, but rather give their best models (that would start from the raw data file and produce the final prediction for each row in the test data), and the judges would then run the models on the test data.\n\nBTW: I apologize if my English is not perfect. I come from the (true) middle-east.", "406999": "Hello Kaggle", "406614": "Thanks for addressing the issue and update. Finally 'real-world' problem! ", "405565": "Sounds like no new columns were added, only removed, is that correct?", "404759": "I was parallely working on two competitions, this one and other in Hackerearth. Both of them ended with data leakage.\nThe Hackerearth test data had 99% data copied from training set , resulting in perfect AUC score of .9999 on leaderboard\nFunny thing is that in spite of bringing it to their notice by few of us, a week ago, they are altogether denying the leak ignoring all the proofs.\nAt least here we have acknowledgement of leak", "404725": "Mark", "404706": "I expect that we will soon receive positive news and a revision of the deadline and rules. ", "404628": "I left the Airbus competition in between (when the test image leak was reported) to join this one hoping not to have same issue here...now that this one also had a leak, I might as well go back to Airbus (that one claims to have fixed the leak) :) Hopefully this competition will provide additional time once the issues is fixed\n", "404535": "Nice I got this update early and hoping for a revision soon.", "404532": "Thanks for the update. I hope you update the new competition soon", "404284": "I am not sure how any of the test data can be salvaged. You would have to remove visitID, which is the timestamp thus a valuable piece of information. You would also have to remove pageviews and hits, which are the most valuable columns in the data. If someone manages to scrape all the session level data from Analytics, they can likely link it to every session in the test data. \nI am concerned that Kaggle will try to create a new dataset that ends up being leaky anyway", "403055": "Thanks for the update, i hope something is done to keep things going smoothly, i was really in the competition spirit.", "402743": "Maybe we can just change the batch of data? Like using the most latest data which answers are not available. ", "402547": "As I understand correctly - the current dataset will be changed?", "413552": "", "405194": "", "405138": "", "404983": "", "404929": "", "404858": "", "403447": "", "403021": "", "402850": "", "402597": "", "402542": "Thank you Julia.", "415388": "Thanks ", "406244": "Thank you for the update", "406234": "thanks for the update", "405177": "Thanks for the update", "404481": "Thank you !!!" } }