Hi All,\nI miss seeing the discussions here alive =)\nHas anyone started modeling towards Part B of the competition?
\n\nI started out using Part A's data, and have a baseline CV score of 0.1897 in predicting the target (new rules).
\n\nCan't wait to get the new data so we can go on with this competition.
", "messages": [ { "id": "409728", "postDate": "10/24/2018 18:56:33", "content": "Hi All,\nI miss seeing the discussions here alive =)\nHas anyone started modeling towards Part B of the competition?
\n\nI started out using Part A's data, and have a baseline CV score of 0.1897 in predicting the target (new rules).
\n\nCan't wait to get the new data so we can go on with this competition.
", "rawMarkdown": "Hi All,\nI miss seeing the discussions here alive =)\nHas anyone started modeling towards Part B of the competition?\n\nI started out using Part A's data, and have a baseline CV score of 0.1897 in predicting the target (new rules).\n\nCan't wait to get the new data so we can go on with this competition.", "votes": null }, { "id": "409737", "postDate": "10/24/2018 19:29:24", "content": "can you explain what do you mean by new rules?\nUsing hits, pageviews we could get 1.59 cv score. But your cv score is 0.1897 (which is really great!!)
\n\ncan you tell more about the method you used?
", "rawMarkdown": "can you explain what do you mean by new rules?\nUsing hits, pageviews we could get 1.59 cv score. But your cv score is 0.1897 (which is really great!!)\n\ncan you tell more about the method you used?", "votes": null }, { "id": "409746", "postDate": "10/24/2018 19:40:52", "content": "Hi @Suchith,\nIn Part A we were asked to predict the total revenue for each visitor in a period in which we have his sessions data.
\n\nIn Part B (that we will have final data for in a couple of weeks) we are asked to predict a visitor's total revenue for a future, unseen period in which we have no sessions yet.
\n\nI used the data from Part A to model target prediction for the new requirement, in order to prepare myself better for when we get the data, the RMSE is much lower in the new requirement, as there are significantly more 0s in the target.
", "rawMarkdown": "Hi @Suchith,\nIn Part A we were asked to predict the total revenue for each visitor in a period in which we have his sessions data.\n\nIn Part B (that we will have final data for in a couple of weeks) we are asked to predict a visitor's total revenue for a future, unseen period in which we have no sessions yet.\n\nI used the data from Part A to model target prediction for the new requirement, in order to prepare myself better for when we get the data, the RMSE is much lower in the new requirement, as there are significantly more 0s in the target.", "votes": null }, { "id": "409876", "postDate": "10/25/2018 02:00:48", "content": "Have you considered seasonal factors？Because we predict that the time will pass through Christmas.
", "rawMarkdown": "Have you considered seasonal factors？Because we predict that the time will pass through Christmas.", "votes": null }, { "id": "409939", "postDate": "10/25/2018 05:20:19", "content": "Hi @guang yang\nNot yet! \nThe old train data is much shorter, so my baseline does not include seasonal features.
", "rawMarkdown": "Hi @guang yang\nNot yet! \nThe old train data is much shorter, so my baseline does not include seasonal features.", "votes": null }, { "id": "409998", "postDate": "10/25/2018 07:16:02", "content": "AmirH, your result sounds really good comparing to original competition. \nAre you prediction both buyers and zero revenue customers? As predicting buyers only is a simpler task.\nAre you prediction for next 3 month interval?
", "rawMarkdown": "AmirH, your result sounds really good comparing to original competition. \nAre you prediction both buyers and zero revenue customers? As predicting buyers only is a simpler task.\nAre you prediction for next 3 month interval?", "votes": null }, { "id": "410005", "postDate": "10/25/2018 07:25:59", "content": "Hi @Alexander Firsov.
\n\nYes, i am predicting both zero revenue and actual buying customers.\nAnd yes, i am predicting a target period which is 3 months after the latest training set date.
\n\nThe reason for the lower score is that there are a lot more zeros vs non zeros in comparison to Part A of this competition.\nThat makes the RMSE significantly lower.\nthe ratio in my CV of buyers\\non buyers is 0.0001%~
", "rawMarkdown": "Hi @Alexander Firsov.\n\nYes, i am predicting both zero revenue and actual buying customers.\nAnd yes, i am predicting a target period which is 3 months after the latest training set date.\n\nThe reason for the lower score is that there are a lot more zeros vs non zeros in comparison to Part A of this competition.\nThat makes the RMSE significantly lower.\nthe ratio in my CV of buyers\\non buyers is 0.0001%~", "votes": null }, { "id": "412159", "postDate": "10/29/2018 17:52:41", "content": "@AmirH
\n\nThat doesn’t sound quite right. That sounds like only 1 in 13,000 of the initial 1.3% of customers becomes a repeat customer (in the window).
\n\nCan you check what score you get with a very simple baseline: just always predict zero (no revenue). I get a RMSE (ie sqrt of mean of log1p(y)^2) of ~2.3.
", "rawMarkdown": "AmirH \n\nThat doesn’t sound quite right. That sounds like only 1 in 13,000 of the initial 1.3% of customers becomes a repeat customer (in the window). \n\nCan you check what score you get with a very simple baseline: just always predict zero (no revenue). I get a RMSE (ie sqrt of mean of log1p(y)^2) of ~2.3.", "votes": null }, { "id": "412176", "postDate": "10/29/2018 18:20:23", "content": "Hi @Burrito Dan,
\n\nI'm not sure, \nlet me clarify what exactly I did:\nI split the time in training set to:
\n\nTraining: Before 20170501
\n\nTarget Prediction Period: After 20170701 (until end of august, that's what we have)
\n\nTotal amount of users in training period that buy in target period: 66
\n\nOut of these 66, only 22 had transactions during the training period.
\n\nmy 0 prediction RMSE for all visitors that appear in training set (559825) : 0.195
\n\nmy 0 prediction RMSE for all visitors that appear in training and also have transactions in training(7292): 1.029
\n\nmy 0.1897 RMSE score is for all visitors that appear in the training set, and for them you need to predict the total transactionRevenue in the target period.
", "rawMarkdown": "Hi @Burrito Dan,\n\nI'm not sure, \nlet me clarify what exactly I did:\nI split the time in training set to:\n\nTraining: Before 20170501\n\nTarget Prediction Period: After 20170701 (until end of august, that's what we have)\n\n\nTotal amount of users in training period that buy in target period: 66\n\nOut of these 66, only 22 had transactions during the training period.\n\nmy 0 prediction RMSE for all visitors that appear in training set (559825) : 0.195\n\nmy 0 prediction RMSE for all visitors that appear in training and also have transactions in training(7292): 1.029\n\n\n\nmy 0.1897 RMSE score is for all visitors that appear in the training set, and for them you need to predict the total transactionRevenue in the target period.", "votes": null } ], "comments": [ { "id": 409737, "author_name": "suchith0312", "author_url": "", "post_date": "10/24/2018 19:29:24", "content": "can you explain what do you mean by new rules?\nUsing hits, pageviews we could get 1.59 cv score. But your cv score is 0.1897 (which is really great!!)
\n\ncan you tell more about the method you used?
", "votes": null, "replies": [ { "id": 409746, "author_name": "tpthegreat", "author_url": "", "post_date": "10/24/2018 19:40:52", "content": "Hi @Suchith,\nIn Part A we were asked to predict the total revenue for each visitor in a period in which we have his sessions data.
\n\nIn Part B (that we will have final data for in a couple of weeks) we are asked to predict a visitor's total revenue for a future, unseen period in which we have no sessions yet.
\n\nI used the data from Part A to model target prediction for the new requirement, in order to prepare myself better for when we get the data, the RMSE is much lower in the new requirement, as there are significantly more 0s in the target.
", "votes": null, "replies": [] } ] }, { "id": 409876, "author_name": "guang10", "author_url": "", "post_date": "10/25/2018 02:00:48", "content": "Have you considered seasonal factors？Because we predict that the time will pass through Christmas.
", "votes": null, "replies": [ { "id": 409939, "author_name": "tpthegreat", "author_url": "", "post_date": "10/25/2018 05:20:19", "content": "Hi @guang yang\nNot yet! \nThe old train data is much shorter, so my baseline does not include seasonal features.
", "votes": null, "replies": [] } ] }, { "id": 409998, "author_name": "alexfir", "author_url": "", "post_date": "10/25/2018 07:16:02", "content": "AmirH, your result sounds really good comparing to original competition. \nAre you prediction both buyers and zero revenue customers? As predicting buyers only is a simpler task.\nAre you prediction for next 3 month interval?
", "votes": null, "replies": [ { "id": 410005, "author_name": "tpthegreat", "author_url": "", "post_date": "10/25/2018 07:25:59", "content": "Hi @Alexander Firsov.
\n\nYes, i am predicting both zero revenue and actual buying customers.\nAnd yes, i am predicting a target period which is 3 months after the latest training set date.
\n\nThe reason for the lower score is that there are a lot more zeros vs non zeros in comparison to Part A of this competition.\nThat makes the RMSE significantly lower.\nthe ratio in my CV of buyers\\non buyers is 0.0001%~
", "votes": null, "replies": [] }, { "id": 412159, "author_name": "burritodan", "author_url": "", "post_date": "10/29/2018 17:52:41", "content": "@AmirH
\n\nThat doesn’t sound quite right. That sounds like only 1 in 13,000 of the initial 1.3% of customers becomes a repeat customer (in the window).
\n\nCan you check what score you get with a very simple baseline: just always predict zero (no revenue). I get a RMSE (ie sqrt of mean of log1p(y)^2) of ~2.3.
", "votes": null, "replies": [] }, { "id": 412176, "author_name": "tpthegreat", "author_url": "", "post_date": "10/29/2018 18:20:23", "content": "Hi @Burrito Dan,
\n\nI'm not sure, \nlet me clarify what exactly I did:\nI split the time in training set to:
\n\nTraining: Before 20170501
\n\nTarget Prediction Period: After 20170701 (until end of august, that's what we have)
\n\nTotal amount of users in training period that buy in target period: 66
\n\nOut of these 66, only 22 had transactions during the training period.
\n\nmy 0 prediction RMSE for all visitors that appear in training set (559825) : 0.195
\n\nmy 0 prediction RMSE for all visitors that appear in training and also have transactions in training(7292): 1.029
\n\nmy 0.1897 RMSE score is for all visitors that appear in the training set, and for them you need to predict the total transactionRevenue in the target period.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "409728": "Hi All,\nI miss seeing the discussions here alive =)\nHas anyone started modeling towards Part B of the competition?\n\nI started out using Part A's data, and have a baseline CV score of 0.1897 in predicting the target (new rules).\n\nCan't wait to get the new data so we can go on with this competition.", "409737": "can you explain what do you mean by new rules?\nUsing hits, pageviews we could get 1.59 cv score. But your cv score is 0.1897 (which is really great!!)\n\ncan you tell more about the method you used?", "409746": "Hi @Suchith,\nIn Part A we were asked to predict the total revenue for each visitor in a period in which we have his sessions data.\n\nIn Part B (that we will have final data for in a couple of weeks) we are asked to predict a visitor's total revenue for a future, unseen period in which we have no sessions yet.\n\nI used the data from Part A to model target prediction for the new requirement, in order to prepare myself better for when we get the data, the RMSE is much lower in the new requirement, as there are significantly more 0s in the target.", "409876": "Have you considered seasonal factors？Because we predict that the time will pass through Christmas.", "409939": "Hi @guang yang\nNot yet! \nThe old train data is much shorter, so my baseline does not include seasonal features.", "409998": "AmirH, your result sounds really good comparing to original competition. \nAre you prediction both buyers and zero revenue customers? As predicting buyers only is a simpler task.\nAre you prediction for next 3 month interval?", "410005": "Hi @Alexander Firsov.\n\nYes, i am predicting both zero revenue and actual buying customers.\nAnd yes, i am predicting a target period which is 3 months after the latest training set date.\n\nThe reason for the lower score is that there are a lot more zeros vs non zeros in comparison to Part A of this competition.\nThat makes the RMSE significantly lower.\nthe ratio in my CV of buyers\\non buyers is 0.0001%~", "412159": "AmirH \n\nThat doesn’t sound quite right. That sounds like only 1 in 13,000 of the initial 1.3% of customers becomes a repeat customer (in the window). \n\nCan you check what score you get with a very simple baseline: just always predict zero (no revenue). I get a RMSE (ie sqrt of mean of log1p(y)^2) of ~2.3.", "412176": "Hi @Burrito Dan,\n\nI'm not sure, \nlet me clarify what exactly I did:\nI split the time in training set to:\n\nTraining: Before 20170501\n\nTarget Prediction Period: After 20170701 (until end of august, that's what we have)\n\n\nTotal amount of users in training period that buy in target period: 66\n\nOut of these 66, only 22 had transactions during the training period.\n\nmy 0 prediction RMSE for all visitors that appear in training set (559825) : 0.195\n\nmy 0 prediction RMSE for all visitors that appear in training and also have transactions in training(7292): 1.029\n\n\n\nmy 0.1897 RMSE score is for all visitors that appear in the training set, and for them you need to predict the total transactionRevenue in the target period." }, "source": "meta" }