Steady 5 submissions per day since joined this competition 6 days ago, score path with 1bps boost daily as follows:
\n0.509 (day 1) -> 0.508 (day 2) -> 0.507 (day 3) -> 0.506 (day 4) -> 0.505 (day 5) -> 0.504 (day 6).
I can't even over-fit the LB with deep parameters...
\nI have to say - Grandmaster definitely has some magic!
Steady 5 submissions per day since joined this competition 6 days ago, score path with 1bps boost daily as follows:
\n0.509 (day 1) -> 0.508 (day 2) -> 0.507 (day 3) -> 0.506 (day 4) -> 0.505 (day 5) -> 0.504 (day 6).
I can't even over-fit the LB with deep parameters...
\nI have to say - Grandmaster definitely has some magic!
Well, that's great!\nDo you mind sharing some insights so that other can learn something?
", "rawMarkdown": "Well, that's great!\nDo you mind sharing some insights so that other can learn something?", "votes": null }, { "id": "267967", "postDate": "01/12/2018 22:49:22", "content": "Hi FabSchreiber, I did nothing fancy, just train my LGBM, XGBoost, CatBoost models offline with fine tuned parameters, mixed training windows, and some more engineered features from the other input files. If I discover magic tricks, I am not reluctant to reveal them.
", "rawMarkdown": "Hi FabSchreiber, I did nothing fancy, just train my LGBM, XGBoost, CatBoost models offline with fine tuned parameters, mixed training windows, and some more engineered features from the other input files. If I discover magic tricks, I am not reluctant to reveal them.", "votes": null }, { "id": "268170", "postDate": "01/13/2018 16:17:51", "content": "Hi FabSchreiber, I think DSEverything is talking about the Kaggle grandmaster Shize Su, who went from .509 to .504 in six days.
", "rawMarkdown": "Hi FabSchreiber, I think DSEverything is talking about the Kaggle grandmaster Shize Su, who went from .509 to .504 in six days.", "votes": null }, { "id": "268817", "postDate": "01/15/2018 18:03:29", "content": "Update: I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603
\n\nThanks for the kind words, DSEverything^_^
\n\nI entered this competition when there is only 10 days left (after finishing the last Master-only competition) with the original goal as simply doing it for some fun in case I have some spare time, but I am happy (and a bit surprised) with the score I achieved so far in this competition.
\n\nI would be happy to share my approach after the competition ends (maybe tomorrow, today is a busy day for me). For now, I would simply say that a somewhat reliable CV mechanism and feature engineering are the key for me to make smooth progress in this competition. My lgb gave me 0.505 on LB, and my nn also gave me 0.505 on the LB, and the ensemble of them gave my final score. I would leave the more detailed information sharing after the competition ends.
\n\nI am all done in this competition. Good luck to everyone!
\n\nBest regards,
\n\nShize
", "rawMarkdown": "Update: I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603\n\n\nThanks for the kind words, DSEverything^_^ \n\nI entered this competition when there is only 10 days left (after finishing the last Master-only competition) with the original goal as simply doing it for some fun in case I have some spare time, but I am happy (and a bit surprised) with the score I achieved so far in this competition.\n\nI would be happy to share my approach after the competition ends (maybe tomorrow, today is a busy day for me). For now, I would simply say that a somewhat reliable CV mechanism and feature engineering are the key for me to make smooth progress in this competition. My lgb gave me 0.505 on LB, and my nn also gave me 0.505 on the LB, and the ensemble of them gave my final score. I would leave the more detailed information sharing after the competition ends.\n\nI am all done in this competition. Good luck to everyone!\n\nBest regards,\n\nShize", "votes": null }, { "id": "268850", "postDate": "01/15/2018 20:02:18", "content": "looking forward to seeing your solution, I'm having trouble with finding a good CV strategy myself :(
", "rawMarkdown": "looking forward to seeing your solution, I'm having trouble with finding a good CV strategy myself :(", "votes": null }, { "id": "269549", "postDate": "01/17/2018 01:28:44", "content": "I honestly don't know how to do a cross validation strategy or any code that will work on a kaggle kernal. I can get basic one's to execute on my computer, but not through kaggle. I done the titanic dataset on kaggle, but beyond that dataset, I have struggled.
\n\nSo I would really appreciate this post too.
", "rawMarkdown": "I honestly don't know how to do a cross validation strategy or any code that will work on a kaggle kernal. I can get basic one's to execute on my computer, but not through kaggle. I done the titanic dataset on kaggle, but beyond that dataset, I have struggled.\n\nSo I would really appreciate this post too.", "votes": null }, { "id": "269551", "postDate": "01/17/2018 01:31:47", "content": "hi steelrose, I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603
", "rawMarkdown": "hi steelrose, I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603", "votes": null }, { "id": "269575", "postDate": "01/17/2018 02:21:11", "content": "There are some question for me too. Do you mind share more details?
\n\nFirst, why the data between 20170720 and 20170725 is not used in cv experiments?
\n\nSecond, as you said \" statistical features were generated independent for each week of the training data, validation data, and test data.\", the training data and test data is the cv experiments? or How to do the cv experiments and avoid the overfitting?
", "rawMarkdown": "There are some question for me too. Do you mind share more details? \n \nFirst, why the data between 20170720 and 20170725 is not used in cv experiments?\n \nSecond, as you said \" statistical features were generated independent for each week of the training data, validation data, and test data.\", the training data and test data is the cv experiments? or How to do the cv experiments and avoid the overfitting?", "votes": null }, { "id": "269598", "postDate": "01/17/2018 02:47:07", "content": "@haifeng:
\n\nnote that for the LB period, it started on Wednesday (2017/08/16), and we need to predict the future 16 days. To make CV experiments more similar to LB, we also generate features for each Wednesday (rather than Monday, Tuesday, Thursday, ...Sunday) to preserve the same DayOfWeek dynamics, and predict the future 16 days. That's why the date 2017/07/20-2017/07/25 were not directly used. However, note that, the sales data in 2017/07/20-2017/07/25 are not ignored, they were used in the historical sales averages and stats for the date 2017/07/26 (and similarly, the sales data for 2017/07/13 -2017/07/18 are not directly used, but only used in computing the features for date 2017/07/19, etc. This is also the framework used in the popular shared lgb starter script.
\n\nAbout generating the statistical features, in order to avoid overfitting to CV, the key is to be careful to avoid utilizing (explicitly or implicitly) any future target information when generating these features. By generating the statistical features for each week independently, we could completely avoid such future target information leak since at that time all the past sales data definitely not including any future sales data information, and the derived statistical features based on these past sales data of course is impossible to contain leak info from future targets.
\n\nFinally, for training data, validation data and test data, I meant my engineered feature set for the period 2017/05/03-2017/07/19, 2017/07/26-2017/08/10, and 2017/08/16-2017/08/31, respectively, as described in my sampled data section.
\n\nBtw, if you want to discuss more, please let's us move to the post https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603, to be concentrated.
", "rawMarkdown": "haifeng: \n\nnote that for the LB period, it started on Wednesday (2017/08/16), and we need to predict the future 16 days. To make CV experiments more similar to LB, we also generate features for each Wednesday (rather than Monday, Tuesday, Thursday, ...Sunday) to preserve the same DayOfWeek dynamics, and predict the future 16 days. That's why the date 2017/07/20-2017/07/25 were not directly used. However, note that, the sales data in 2017/07/20-2017/07/25 are not ignored, they were used in the historical sales averages and stats for the date 2017/07/26 (and similarly, the sales data for 2017/07/13 -2017/07/18 are not directly used, but only used in computing the features for date 2017/07/19, etc. This is also the framework used in the popular shared lgb starter script.\n\nAbout generating the statistical features, in order to avoid overfitting to CV, the key is to be careful to avoid utilizing (explicitly or implicitly) any future target information when generating these features. By generating the statistical features for each week independently, we could completely avoid such future target information leak since at that time all the past sales data definitely not including any future sales data information, and the derived statistical features based on these past sales data of course is impossible to contain leak info from future targets.\n\nFinally, for training data, validation data and test data, I meant my engineered feature set for the period 2017/05/03-2017/07/19, 2017/07/26-2017/08/10, and 2017/08/16-2017/08/31, respectively, as described in my sampled data section.\n\nBtw, if you want to discuss more, please let's us move to the post https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603, to be concentrated.", "votes": null }, { "id": "269654", "postDate": "01/17/2018 04:16:10", "content": "Thanks! It's very helpful!!!
", "rawMarkdown": "Thanks! It's very helpful!!!", "votes": null } ], "comments": [ { "id": 267761, "author_name": "fabsta", "author_url": "", "post_date": "01/12/2018 07:59:19", "content": "Well, that's great!\nDo you mind sharing some insights so that other can learn something?
", "votes": null, "replies": [ { "id": 267967, "author_name": "dongxu027", "author_url": "", "post_date": "01/12/2018 22:49:22", "content": "Hi FabSchreiber, I did nothing fancy, just train my LGBM, XGBoost, CatBoost models offline with fine tuned parameters, mixed training windows, and some more engineered features from the other input files. If I discover magic tricks, I am not reluctant to reveal them.
", "votes": null, "replies": [] }, { "id": 268170, "author_name": "clark78118", "author_url": "", "post_date": "01/13/2018 16:17:51", "content": "Hi FabSchreiber, I think DSEverything is talking about the Kaggle grandmaster Shize Su, who went from .509 to .504 in six days.
", "votes": null, "replies": [] } ] }, { "id": 268817, "author_name": "sushize", "author_url": "", "post_date": "01/15/2018 18:03:29", "content": "Update: I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603
\n\nThanks for the kind words, DSEverything^_^
\n\nI entered this competition when there is only 10 days left (after finishing the last Master-only competition) with the original goal as simply doing it for some fun in case I have some spare time, but I am happy (and a bit surprised) with the score I achieved so far in this competition.
\n\nI would be happy to share my approach after the competition ends (maybe tomorrow, today is a busy day for me). For now, I would simply say that a somewhat reliable CV mechanism and feature engineering are the key for me to make smooth progress in this competition. My lgb gave me 0.505 on LB, and my nn also gave me 0.505 on the LB, and the ensemble of them gave my final score. I would leave the more detailed information sharing after the competition ends.
\n\nI am all done in this competition. Good luck to everyone!
\n\nBest regards,
\n\nShize
", "votes": null, "replies": [ { "id": 268850, "author_name": "steelrose", "author_url": "", "post_date": "01/15/2018 20:02:18", "content": "looking forward to seeing your solution, I'm having trouble with finding a good CV strategy myself :(
", "votes": null, "replies": [] }, { "id": 269549, "author_name": "skyrmion", "author_url": "", "post_date": "01/17/2018 01:28:44", "content": "I honestly don't know how to do a cross validation strategy or any code that will work on a kaggle kernal. I can get basic one's to execute on my computer, but not through kaggle. I done the titanic dataset on kaggle, but beyond that dataset, I have struggled.
\n\nSo I would really appreciate this post too.
", "votes": null, "replies": [] }, { "id": 269551, "author_name": "sushize", "author_url": "", "post_date": "01/17/2018 01:31:47", "content": "hi steelrose, I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603
", "votes": null, "replies": [] }, { "id": 269575, "author_name": "hubifeng", "author_url": "", "post_date": "01/17/2018 02:21:11", "content": "There are some question for me too. Do you mind share more details?
\n\nFirst, why the data between 20170720 and 20170725 is not used in cv experiments?
\n\nSecond, as you said \" statistical features were generated independent for each week of the training data, validation data, and test data.\", the training data and test data is the cv experiments? or How to do the cv experiments and avoid the overfitting?
", "votes": null, "replies": [] }, { "id": 269598, "author_name": "sushize", "author_url": "", "post_date": "01/17/2018 02:47:07", "content": "@haifeng:
\n\nnote that for the LB period, it started on Wednesday (2017/08/16), and we need to predict the future 16 days. To make CV experiments more similar to LB, we also generate features for each Wednesday (rather than Monday, Tuesday, Thursday, ...Sunday) to preserve the same DayOfWeek dynamics, and predict the future 16 days. That's why the date 2017/07/20-2017/07/25 were not directly used. However, note that, the sales data in 2017/07/20-2017/07/25 are not ignored, they were used in the historical sales averages and stats for the date 2017/07/26 (and similarly, the sales data for 2017/07/13 -2017/07/18 are not directly used, but only used in computing the features for date 2017/07/19, etc. This is also the framework used in the popular shared lgb starter script.
\n\nAbout generating the statistical features, in order to avoid overfitting to CV, the key is to be careful to avoid utilizing (explicitly or implicitly) any future target information when generating these features. By generating the statistical features for each week independently, we could completely avoid such future target information leak since at that time all the past sales data definitely not including any future sales data information, and the derived statistical features based on these past sales data of course is impossible to contain leak info from future targets.
\n\nFinally, for training data, validation data and test data, I meant my engineered feature set for the period 2017/05/03-2017/07/19, 2017/07/26-2017/08/10, and 2017/08/16-2017/08/31, respectively, as described in my sampled data section.
\n\nBtw, if you want to discuss more, please let's us move to the post https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603, to be concentrated.
", "votes": null, "replies": [] }, { "id": 269654, "author_name": "hubifeng", "author_url": "", "post_date": "01/17/2018 04:16:10", "content": "Thanks! It's very helpful!!!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "267683": "Steady 5 submissions per day since joined this competition 6 days ago, score path with 1bps boost daily as follows: \n0.509 (day 1) -> 0.508 (day 2) -> 0.507 (day 3) -> 0.506 (day 4) -> 0.505 (day 5) -> 0.504 (day 6). \n \nI can't even over-fit the LB with deep parameters... \nI have to say - Grandmaster definitely has some magic!", "267761": "Well, that's great!\nDo you mind sharing some insights so that other can learn something?", "267967": "Hi FabSchreiber, I did nothing fancy, just train my LGBM, XGBoost, CatBoost models offline with fine tuned parameters, mixed training windows, and some more engineered features from the other input files. If I discover magic tricks, I am not reluctant to reveal them.", "268170": "Hi FabSchreiber, I think DSEverything is talking about the Kaggle grandmaster Shize Su, who went from .509 to .504 in six days.", "268817": "Update: I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603\n\n\nThanks for the kind words, DSEverything^_^ \n\nI entered this competition when there is only 10 days left (after finishing the last Master-only competition) with the original goal as simply doing it for some fun in case I have some spare time, but I am happy (and a bit surprised) with the score I achieved so far in this competition.\n\nI would be happy to share my approach after the competition ends (maybe tomorrow, today is a busy day for me). For now, I would simply say that a somewhat reliable CV mechanism and feature engineering are the key for me to make smooth progress in this competition. My lgb gave me 0.505 on LB, and my nn also gave me 0.505 on the LB, and the ensemble of them gave my final score. I would leave the more detailed information sharing after the competition ends.\n\nI am all done in this competition. Good luck to everyone!\n\nBest regards,\n\nShize", "268850": "looking forward to seeing your solution, I'm having trouble with finding a good CV strategy myself :(", "269549": "I honestly don't know how to do a cross validation strategy or any code that will work on a kaggle kernal. I can get basic one's to execute on my computer, but not through kaggle. I done the titanic dataset on kaggle, but beyond that dataset, I have struggled.\n\nSo I would really appreciate this post too.", "269551": "hi steelrose, I have already shared my approach here https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603", "269575": "There are some question for me too. Do you mind share more details? \n \nFirst, why the data between 20170720 and 20170725 is not used in cv experiments?\n \nSecond, as you said \" statistical features were generated independent for each week of the training data, validation data, and test data.\", the training data and test data is the cv experiments? or How to do the cv experiments and avoid the overfitting?", "269598": "haifeng: \n\nnote that for the LB period, it started on Wednesday (2017/08/16), and we need to predict the future 16 days. To make CV experiments more similar to LB, we also generate features for each Wednesday (rather than Monday, Tuesday, Thursday, ...Sunday) to preserve the same DayOfWeek dynamics, and predict the future 16 days. That's why the date 2017/07/20-2017/07/25 were not directly used. However, note that, the sales data in 2017/07/20-2017/07/25 are not ignored, they were used in the historical sales averages and stats for the date 2017/07/26 (and similarly, the sales data for 2017/07/13 -2017/07/18 are not directly used, but only used in computing the features for date 2017/07/19, etc. This is also the framework used in the popular shared lgb starter script.\n\nAbout generating the statistical features, in order to avoid overfitting to CV, the key is to be careful to avoid utilizing (explicitly or implicitly) any future target information when generating these features. By generating the statistical features for each week independently, we could completely avoid such future target information leak since at that time all the past sales data definitely not including any future sales data information, and the derived statistical features based on these past sales data of course is impossible to contain leak info from future targets.\n\nFinally, for training data, validation data and test data, I meant my engineered feature set for the period 2017/05/03-2017/07/19, 2017/07/26-2017/08/10, and 2017/08/16-2017/08/31, respectively, as described in my sampled data section.\n\nBtw, if you want to discuss more, please let's us move to the post https://www.kaggle.com/c/favorita-grocery-sales-forecasting/discussion/47603, to be concentrated.", "269654": "Thanks! It's very helpful!!!" }, "source": "meta" }