Hi Kagglers,\nCould you maybe share about how each method worked for you?
\n\nSession based - simple log1p(sum(predictions)) by fullVisitorId
User based
Session based -> User Based - train session based and use predictions to train another User based model
And where inside those options did Classification give you a boost?
\n\nI'm still a bit in the dark in this competition so i don't have a lot to share.\nMy best model was session based at about 1.44~
\n\nThanks!!
", "messages": [ { "id": "396680", "postDate": "10/01/2018 07:04:58", "content": "Hi Kagglers,\nCould you maybe share about how each method worked for you?
\n\nSession based - simple log1p(sum(predictions)) by fullVisitorId
User based
Session based -> User Based - train session based and use predictions to train another User based model
And where inside those options did Classification give you a boost?
\n\nI'm still a bit in the dark in this competition so i don't have a lot to share.\nMy best model was session based at about 1.44~
\n\nThanks!!
", "rawMarkdown": "Hi Kagglers,\nCould you maybe share about how each method worked for you?\n\n- Session based - simple log1p(sum(predictions)) by fullVisitorId\n\n- User based \n\n- Session based -> User Based - train session based and use predictions to train another User based model \n\n\nAnd where inside those options did Classification give you a boost?\n\nI'm still a bit in the dark in this competition so i don't have a lot to share.\nMy best model was session based at about 1.44~\n\nThanks!!", "votes": null }, { "id": "396741", "postDate": "10/01/2018 09:57:34", "content": "Session based: 1.4409\nSession->user: 1.4302
\n\nMy current LB score is a blend of my best submission with the top public kernel.
", "rawMarkdown": "Session based: 1.4409\nSession->user: 1.4302\n\nMy current LB score is a blend of my best submission with the top public kernel.", "votes": null }, { "id": "396759", "postDate": "10/01/2018 10:18:14", "content": "Xgboost session based: 1.4489
", "rawMarkdown": "Xgboost session based: 1.4489", "votes": null }, { "id": "397141", "postDate": "10/02/2018 01:02:58", "content": "Session based LightGBM: 1.4375
", "rawMarkdown": "Session based LightGBM: 1.4375", "votes": null }, { "id": "397238", "postDate": "10/02/2018 07:02:36", "content": "Single lgb model scores as follow:\n[Session-based 1.4362]\n[User-based 1.4314]\n[Session-based -> User-Based 1.4241]
", "rawMarkdown": "Single lgb model scores as follow:\n[Session-based 1.4362]\n[User-based 1.4314]\n[Session-based -> User-Based 1.4241]", "votes": null }, { "id": "397270", "postDate": "10/02/2018 08:10:23", "content": "How much was your score on the CV data?
", "rawMarkdown": "How much was your score on the CV data?", "votes": null }, { "id": "397362", "postDate": "10/02/2018 11:52:43", "content": "New Single LGBM session based - 1.4285
", "rawMarkdown": "New Single LGBM session based - 1.4285", "votes": null }, { "id": "397486", "postDate": "10/02/2018 14:42:35", "content": "Session- or visitor-based?
", "rawMarkdown": "Session- or visitor-based?", "votes": null }, { "id": "397508", "postDate": "10/02/2018 15:11:51", "content": "What did you change? Was it the bug there or more advanced feature engineering?
", "rawMarkdown": "What did you change? Was it the bug there or more advanced feature engineering?", "votes": null }, { "id": "397532", "postDate": "10/02/2018 16:17:52", "content": "Hi leonid.\nSession based
", "rawMarkdown": "Hi leonid.\nSession based", "votes": null }, { "id": "397534", "postDate": "10/02/2018 16:19:18", "content": "Hi Stanislav, there was a bug that i still don't know what it was, fixing it brought me to 1.44.\nFE brought me to 1.428
", "rawMarkdown": "Hi Stanislav, there was a bug that i still don't know what it was, fixing it brought me to 1.44.\nFE brought me to 1.428", "votes": null }, { "id": "397577", "postDate": "10/02/2018 17:41:03", "content": "I'm also using R, and I have the same situation and after complex feature engineering score is upper that 1.45 :)
", "rawMarkdown": "I'm also using R, and I have the same situation and after complex feature engineering score is upper that 1.45 :)", "votes": null }, { "id": "397599", "postDate": "10/02/2018 18:29:54", "content": "Hi @Stanislav,\nOne of the things i did was use the flat data created by @Olivier, don't know if that is what made the difference, but i suggest that you try.
\n\nhttps://www.kaggle.com/ogrellier/create-extracted-json-fields-dataset
\n\ndownload the files here and use them instead of the input file.\nSimple LGBM with this data with added 3 features of \"day of month\",\"day of week\" , \"hour of day\" based on the date in \"visitstarttime\" should give you 1.44~ score.
\n\nThat was my validation that nothing is wrong, i did my FE on that model and achieved my 1.428 score.
", "rawMarkdown": "Hi @Stanislav,\nOne of the things i did was use the flat data created by @Olivier, don't know if that is what made the difference, but i suggest that you try.\n\nhttps://www.kaggle.com/ogrellier/create-extracted-json-fields-dataset\n\ndownload the files here and use them instead of the input file.\nSimple LGBM with this data with added 3 features of \"day of month\",\"day of week\" , \"hour of day\" based on the date in \"visitstarttime\" should give you 1.44~ score.\n\nThat was my validation that nothing is wrong, i did my FE on that model and achieved my 1.428 score.", "votes": null }, { "id": "397776", "postDate": "10/03/2018 05:09:21", "content": "Session->user: 1.4242
", "rawMarkdown": "Session->user: 1.4242", "votes": null }, { "id": "398123", "postDate": "10/03/2018 14:35:57", "content": "Thank you so much!!\nI tried to use new data, that you wrote, but the score became the same - 1.457, where my cv and oof was pretty good\nAfter that, I remove features like year, week number and other date features except for wday and mday and my score became 1.440..but my cv and oof scores became worse on 0.01\nSo it wasn't a bug, I will think about simple features that can lead to overfitting...
", "rawMarkdown": "Thank you so much!!\nI tried to use new data, that you wrote, but the score became the same - 1.457, where my cv and oof was pretty good\nAfter that, I remove features like year, week number and other date features except for wday and mday and my score became 1.440..but my cv and oof scores became worse on 0.01\nSo it wasn't a bug, I will think about simple features that can lead to overfitting...", "votes": null }, { "id": "398186", "postDate": "10/03/2018 16:26:24", "content": "Great to hear @stanislav!\nAfter removing these features and using non over fitting ones, my cv and lb correspond well with one another. \nKeep up the great work!
", "rawMarkdown": "Great to hear @stanislav!\nAfter removing these features and using non over fitting ones, my cv and lb correspond well with one another. \nKeep up the great work!", "votes": null }, { "id": "404777", "postDate": "10/16/2018 11:25:38", "content": "Hello @tpthegreat, do you confirm you reach 1.4285 with a session-level model, i.e. without any user-level aggregated feature?
\n\nThat's pretty remarkable, I can only get to 1.4360 at session-level, in line with the other scores presented here. Do you have some killer features that haven't made their way into public kernels yet?
", "rawMarkdown": "Hello @tpthegreat, do you confirm you reach 1.4285 with a session-level model, i.e. without any user-level aggregated feature? \n\nThat's pretty remarkable, I can only get to 1.4360 at session-level, in line with the other scores presented here. Do you have some killer features that haven't made their way into public kernels yet?", "votes": null }, { "id": "404868", "postDate": "10/16/2018 13:40:41", "content": "Hi @juliencs,\nMaybe so :)\nMy best single lgbm session based is actually 1.4238
", "rawMarkdown": "Hi @juliencs,\nMaybe so :)\nMy best single lgbm session based is actually 1.4238", "votes": null }, { "id": "404871", "postDate": "10/16/2018 13:43:56", "content": "My best single LGBM was giving 1.4161. It was session based.
", "rawMarkdown": "My best single LGBM was giving 1.4161. It was session based.", "votes": null }, { "id": "404904", "postDate": "10/16/2018 14:46:12", "content": "I only got around to doing session-based models before this competition went off the rails. My best one scored 1.4152.
", "rawMarkdown": "I only got around to doing session-based models before this competition went off the rails. My best one scored 1.4152.", "votes": null }, { "id": "404928", "postDate": "10/16/2018 15:25:48", "content": "Hi Dmitriy. May I ask if you used LGBM for that model or NN? It's no problem if you don't disclose.
", "rawMarkdown": "Hi Dmitriy. May I ask if you used LGBM for that model or NN? It's no problem if you don't disclose.", "votes": null }, { "id": "404979", "postDate": "10/16/2018 17:27:40", "content": "That's Awesome @fatihozturk!!!\ndid you use the whole categorical modifications that are in the kernels right now? and combined pairs of categories? if so - did these transformations make a big impact on the score?
\n\nThanks!\nKeep on rocking!
", "rawMarkdown": "That's Awesome @fatihozturk!!!\ndid you use the whole categorical modifications that are in the kernels right now? and combined pairs of categories? if so - did these transformations make a big impact on the score?\n\nThanks!\nKeep on rocking!", "votes": null }, { "id": "404991", "postDate": "10/16/2018 17:52:23", "content": "@fatihozturk I used XGBoost exclusively.
", "rawMarkdown": "fatihozturk I used XGBoost exclusively.", "votes": null } ], "comments": [ { "id": 396741, "author_name": "konradb", "author_url": "", "post_date": "10/01/2018 09:57:34", "content": "Session based: 1.4409\nSession->user: 1.4302
\n\nMy current LB score is a blend of my best submission with the top public kernel.
", "votes": null, "replies": [] }, { "id": 396759, "author_name": "stsopov", "author_url": "", "post_date": "10/01/2018 10:18:14", "content": "Xgboost session based: 1.4489
", "votes": null, "replies": [] }, { "id": 397141, "author_name": "thawatt", "author_url": "", "post_date": "10/02/2018 01:02:58", "content": "Session based LightGBM: 1.4375
", "votes": null, "replies": [] }, { "id": 397238, "author_name": "marcuslin", "author_url": "", "post_date": "10/02/2018 07:02:36", "content": "Single lgb model scores as follow:\n[Session-based 1.4362]\n[User-based 1.4314]\n[Session-based -> User-Based 1.4241]
", "votes": null, "replies": [ { "id": 397270, "author_name": "stsopov", "author_url": "", "post_date": "10/02/2018 08:10:23", "content": "How much was your score on the CV data?
", "votes": null, "replies": [] } ] }, { "id": 397362, "author_name": "tpthegreat", "author_url": "", "post_date": "10/02/2018 11:52:43", "content": "New Single LGBM session based - 1.4285
", "votes": null, "replies": [ { "id": 397486, "author_name": "leonidg", "author_url": "", "post_date": "10/02/2018 14:42:35", "content": "Session- or visitor-based?
", "votes": null, "replies": [] }, { "id": 397508, "author_name": "stsopov", "author_url": "", "post_date": "10/02/2018 15:11:51", "content": "What did you change? Was it the bug there or more advanced feature engineering?
", "votes": null, "replies": [] }, { "id": 397532, "author_name": "tpthegreat", "author_url": "", "post_date": "10/02/2018 16:17:52", "content": "Hi leonid.\nSession based
", "votes": null, "replies": [] }, { "id": 397534, "author_name": "tpthegreat", "author_url": "", "post_date": "10/02/2018 16:19:18", "content": "Hi Stanislav, there was a bug that i still don't know what it was, fixing it brought me to 1.44.\nFE brought me to 1.428
", "votes": null, "replies": [] }, { "id": 397577, "author_name": "stsopov", "author_url": "", "post_date": "10/02/2018 17:41:03", "content": "I'm also using R, and I have the same situation and after complex feature engineering score is upper that 1.45 :)
", "votes": null, "replies": [] }, { "id": 397599, "author_name": "tpthegreat", "author_url": "", "post_date": "10/02/2018 18:29:54", "content": "Hi @Stanislav,\nOne of the things i did was use the flat data created by @Olivier, don't know if that is what made the difference, but i suggest that you try.
\n\nhttps://www.kaggle.com/ogrellier/create-extracted-json-fields-dataset
\n\ndownload the files here and use them instead of the input file.\nSimple LGBM with this data with added 3 features of \"day of month\",\"day of week\" , \"hour of day\" based on the date in \"visitstarttime\" should give you 1.44~ score.
\n\nThat was my validation that nothing is wrong, i did my FE on that model and achieved my 1.428 score.
", "votes": null, "replies": [] }, { "id": 398123, "author_name": "stsopov", "author_url": "", "post_date": "10/03/2018 14:35:57", "content": "Thank you so much!!\nI tried to use new data, that you wrote, but the score became the same - 1.457, where my cv and oof was pretty good\nAfter that, I remove features like year, week number and other date features except for wday and mday and my score became 1.440..but my cv and oof scores became worse on 0.01\nSo it wasn't a bug, I will think about simple features that can lead to overfitting...
", "votes": null, "replies": [] }, { "id": 398186, "author_name": "tpthegreat", "author_url": "", "post_date": "10/03/2018 16:26:24", "content": "Great to hear @stanislav!\nAfter removing these features and using non over fitting ones, my cv and lb correspond well with one another. \nKeep up the great work!
", "votes": null, "replies": [] }, { "id": 404777, "author_name": "juliencs", "author_url": "", "post_date": "10/16/2018 11:25:38", "content": "Hello @tpthegreat, do you confirm you reach 1.4285 with a session-level model, i.e. without any user-level aggregated feature?
\n\nThat's pretty remarkable, I can only get to 1.4360 at session-level, in line with the other scores presented here. Do you have some killer features that haven't made their way into public kernels yet?
", "votes": null, "replies": [] }, { "id": 404868, "author_name": "tpthegreat", "author_url": "", "post_date": "10/16/2018 13:40:41", "content": "Hi @juliencs,\nMaybe so :)\nMy best single lgbm session based is actually 1.4238
", "votes": null, "replies": [] } ] }, { "id": 397776, "author_name": "chocozzz", "author_url": "", "post_date": "10/03/2018 05:09:21", "content": "Session->user: 1.4242
", "votes": null, "replies": [] }, { "id": 404871, "author_name": "fatihozturk", "author_url": "", "post_date": "10/16/2018 13:43:56", "content": "My best single LGBM was giving 1.4161. It was session based.
", "votes": null, "replies": [ { "id": 404979, "author_name": "tpthegreat", "author_url": "", "post_date": "10/16/2018 17:27:40", "content": "That's Awesome @fatihozturk!!!\ndid you use the whole categorical modifications that are in the kernels right now? and combined pairs of categories? if so - did these transformations make a big impact on the score?
\n\nThanks!\nKeep on rocking!
", "votes": null, "replies": [] } ] }, { "id": 404904, "author_name": "dmitriyguller", "author_url": "", "post_date": "10/16/2018 14:46:12", "content": "I only got around to doing session-based models before this competition went off the rails. My best one scored 1.4152.
", "votes": null, "replies": [ { "id": 404928, "author_name": "fatihozturk", "author_url": "", "post_date": "10/16/2018 15:25:48", "content": "Hi Dmitriy. May I ask if you used LGBM for that model or NN? It's no problem if you don't disclose.
", "votes": null, "replies": [] }, { "id": 404991, "author_name": "dmitriyguller", "author_url": "", "post_date": "10/16/2018 17:52:23", "content": "@fatihozturk I used XGBoost exclusively.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "396680": "Hi Kagglers,\nCould you maybe share about how each method worked for you?\n\n- Session based - simple log1p(sum(predictions)) by fullVisitorId\n\n- User based \n\n- Session based -> User Based - train session based and use predictions to train another User based model \n\n\nAnd where inside those options did Classification give you a boost?\n\nI'm still a bit in the dark in this competition so i don't have a lot to share.\nMy best model was session based at about 1.44~\n\nThanks!!", "396741": "Session based: 1.4409\nSession->user: 1.4302\n\nMy current LB score is a blend of my best submission with the top public kernel.", "396759": "Xgboost session based: 1.4489", "397141": "Session based LightGBM: 1.4375", "397238": "Single lgb model scores as follow:\n[Session-based 1.4362]\n[User-based 1.4314]\n[Session-based -> User-Based 1.4241]", "397270": "How much was your score on the CV data?", "397362": "New Single LGBM session based - 1.4285", "397486": "Session- or visitor-based?", "397508": "What did you change? Was it the bug there or more advanced feature engineering?", "397532": "Hi leonid.\nSession based", "397534": "Hi Stanislav, there was a bug that i still don't know what it was, fixing it brought me to 1.44.\nFE brought me to 1.428", "397577": "I'm also using R, and I have the same situation and after complex feature engineering score is upper that 1.45 :)", "397599": "Hi @Stanislav,\nOne of the things i did was use the flat data created by @Olivier, don't know if that is what made the difference, but i suggest that you try.\n\nhttps://www.kaggle.com/ogrellier/create-extracted-json-fields-dataset\n\ndownload the files here and use them instead of the input file.\nSimple LGBM with this data with added 3 features of \"day of month\",\"day of week\" , \"hour of day\" based on the date in \"visitstarttime\" should give you 1.44~ score.\n\nThat was my validation that nothing is wrong, i did my FE on that model and achieved my 1.428 score.", "397776": "Session->user: 1.4242", "398123": "Thank you so much!!\nI tried to use new data, that you wrote, but the score became the same - 1.457, where my cv and oof was pretty good\nAfter that, I remove features like year, week number and other date features except for wday and mday and my score became 1.440..but my cv and oof scores became worse on 0.01\nSo it wasn't a bug, I will think about simple features that can lead to overfitting...", "398186": "Great to hear @stanislav!\nAfter removing these features and using non over fitting ones, my cv and lb correspond well with one another. \nKeep up the great work!", "404777": "Hello @tpthegreat, do you confirm you reach 1.4285 with a session-level model, i.e. without any user-level aggregated feature? \n\nThat's pretty remarkable, I can only get to 1.4360 at session-level, in line with the other scores presented here. Do you have some killer features that haven't made their way into public kernels yet?", "404868": "Hi @juliencs,\nMaybe so :)\nMy best single lgbm session based is actually 1.4238", "404871": "My best single LGBM was giving 1.4161. It was session based.", "404904": "I only got around to doing session-based models before this competition went off the rails. My best one scored 1.4152.", "404928": "Hi Dmitriy. May I ask if you used LGBM for that model or NN? It's no problem if you don't disclose.", "404979": "That's Awesome @fatihozturk!!!\ndid you use the whole categorical modifications that are in the kernels right now? and combined pairs of categories? if so - did these transformations make a big impact on the score?\n\nThanks!\nKeep on rocking!", "404991": "fatihozturk I used XGBoost exclusively." }, "source": "meta" }