Hi ,\nAs this competition with only 6 days to go, how to detect the outliers seems to be the biggest problem to improve the prediction.\nAlthough I tried my best to detect the outliers, my cv AUC score is 0.91XXX and BinaryLogLoss 0.0427 (5 fold , StratifiedKFold). Considering the outliers only account for 1.6% of the training data, the score did not perform well.
\n\nWith detection model, my current single model CV is 3.647 but the LB is only 3.688.
\n\nI would like to ask what is the AUC score for kagglers who also try to detect the outliers in this competition ?
", "messages": [ { "id": "475841", "postDate": "02/21/2019 09:12:51", "content": "Hi ,\nAs this competition with only 6 days to go, how to detect the outliers seems to be the biggest problem to improve the prediction.\nAlthough I tried my best to detect the outliers, my cv AUC score is 0.91XXX and BinaryLogLoss 0.0427 (5 fold , StratifiedKFold). Considering the outliers only account for 1.6% of the training data, the score did not perform well.
\n\nWith detection model, my current single model CV is 3.647 but the LB is only 3.688.
\n\nI would like to ask what is the AUC score for kagglers who also try to detect the outliers in this competition ?
", "rawMarkdown": "Hi ,\nAs this competition with only 6 days to go, how to detect the outliers seems to be the biggest problem to improve the prediction.\nAlthough I tried my best to detect the outliers, my cv AUC score is 0.91XXX and BinaryLogLoss 0.0427 (5 fold , StratifiedKFold). Considering the outliers only account for 1.6% of the training data, the score did not perform well. \n\nWith detection model, my current single model CV is 3.647 but the LB is only 3.688.\n\nI would like to ask what is the AUC score for kagglers who also try to detect the outliers in this competition ?", "votes": null }, { "id": "475854", "postDate": "02/21/2019 09:36:12", "content": "Good job!!\nMy current classifier gives 0.907 AUC and 0.0439 logloss
", "rawMarkdown": "Good job!!\nMy current classifier gives 0.907 AUC and 0.0439 logloss", "votes": null }, { "id": "475891", "postDate": "02/21/2019 10:41:41", "content": "@AmirH Thanks! \nGreat job to your prediction! My LB score is not even close to yours haha.\nI still can not make the gap between CV and LB smaller.\nSo according to your experience, do you think the classifier really help?\nAs a data science beginner, Could I ask what's your model's CV score ? And how do you use the result of your classifier ? \nI used the idea of this good kernel \n(https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier)\nThanks!
", "rawMarkdown": "AmirH Thanks! \nGreat job to your prediction! My LB score is not even close to yours haha.\nI still can not make the gap between CV and LB smaller.\nSo according to your experience, do you think the classifier really help?\nAs a data science beginner, Could I ask what's your model's CV score ? And how do you use the result of your classifier ? \nI used the idea of this good kernel \n(https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier)\nThanks!", "votes": null }, { "id": "476143", "postDate": "02/21/2019 16:50:50", "content": "If you use typical binary classification model and improve the AUC, I do not think it help anything. Must be some out of box thinking that matters.
", "rawMarkdown": "If you use typical binary classification model and improve the AUC, I do not think it help anything. Must be some out of box thinking that matters.", "votes": null }, { "id": "476327", "postDate": "02/22/2019 01:34:40", "content": "@xiaojiu1414\nThank you for the suggestion!\nSo, according to your opinion, outlier detection still works but should be the method different from binary classification? Could l PM you for further suggestion ? (purely intellectual curiosity )
", "rawMarkdown": "xiaojiu1414\nThank you for the suggestion!\nSo, according to your opinion, outlier detection still works but should be the method different from binary classification? Could l PM you for further suggestion ? (purely intellectual curiosity )", "votes": null }, { "id": "477261", "postDate": "02/24/2019 08:48:46", "content": "Hi @C.Y.W\nSorry for not responding sooner.\nI think the classifier does help,\nOur CV for current LB of 3.668 is about 3.629
\n\nWe currently do not use any post processing.\nClassification just adds value to our ensemble.
\n\nKeep up the good work!
", "rawMarkdown": "Hi @C.Y.W\nSorry for not responding sooner.\nI think the classifier does help,\nOur CV for current LB of 3.668 is about 3.629\n\nWe currently do not use any post processing.\nClassification just adds value to our ensemble.\n\nKeep up the good work!", "votes": null }, { "id": "478437", "postDate": "02/26/2019 06:37:05", "content": "Hi @AmirH,\nI did not use any post processing either in my current solution. And thank you for the reply! \nGood luck to your team!
", "rawMarkdown": "Hi @AmirH,\nI did not use any post processing either in my current solution. And thank you for the reply! \nGood luck to your team!", "votes": null } ], "comments": [ { "id": 475854, "author_name": "tpthegreat", "author_url": "", "post_date": "02/21/2019 09:36:12", "content": "Good job!!\nMy current classifier gives 0.907 AUC and 0.0439 logloss
", "votes": null, "replies": [ { "id": 475891, "author_name": "cywwayne", "author_url": "", "post_date": "02/21/2019 10:41:41", "content": "@AmirH Thanks! \nGreat job to your prediction! My LB score is not even close to yours haha.\nI still can not make the gap between CV and LB smaller.\nSo according to your experience, do you think the classifier really help?\nAs a data science beginner, Could I ask what's your model's CV score ? And how do you use the result of your classifier ? \nI used the idea of this good kernel \n(https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier)\nThanks!
", "votes": null, "replies": [] }, { "id": 477261, "author_name": "tpthegreat", "author_url": "", "post_date": "02/24/2019 08:48:46", "content": "Hi @C.Y.W\nSorry for not responding sooner.\nI think the classifier does help,\nOur CV for current LB of 3.668 is about 3.629
\n\nWe currently do not use any post processing.\nClassification just adds value to our ensemble.
\n\nKeep up the good work!
", "votes": null, "replies": [] }, { "id": 478437, "author_name": "cywwayne", "author_url": "", "post_date": "02/26/2019 06:37:05", "content": "Hi @AmirH,\nI did not use any post processing either in my current solution. And thank you for the reply! \nGood luck to your team!
", "votes": null, "replies": [] } ] }, { "id": 476143, "author_name": "xiaojiu1414", "author_url": "", "post_date": "02/21/2019 16:50:50", "content": "If you use typical binary classification model and improve the AUC, I do not think it help anything. Must be some out of box thinking that matters.
", "votes": null, "replies": [ { "id": 476327, "author_name": "cywwayne", "author_url": "", "post_date": "02/22/2019 01:34:40", "content": "@xiaojiu1414\nThank you for the suggestion!\nSo, according to your opinion, outlier detection still works but should be the method different from binary classification? Could l PM you for further suggestion ? (purely intellectual curiosity )
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "475841": "Hi ,\nAs this competition with only 6 days to go, how to detect the outliers seems to be the biggest problem to improve the prediction.\nAlthough I tried my best to detect the outliers, my cv AUC score is 0.91XXX and BinaryLogLoss 0.0427 (5 fold , StratifiedKFold). Considering the outliers only account for 1.6% of the training data, the score did not perform well. \n\nWith detection model, my current single model CV is 3.647 but the LB is only 3.688.\n\nI would like to ask what is the AUC score for kagglers who also try to detect the outliers in this competition ?", "475854": "Good job!!\nMy current classifier gives 0.907 AUC and 0.0439 logloss", "475891": "AmirH Thanks! \nGreat job to your prediction! My LB score is not even close to yours haha.\nI still can not make the gap between CV and LB smaller.\nSo according to your experience, do you think the classifier really help?\nAs a data science beginner, Could I ask what's your model's CV score ? And how do you use the result of your classifier ? \nI used the idea of this good kernel \n(https://www.kaggle.com/waitingli/combining-your-model-with-a-model-without-outlier)\nThanks!", "476143": "If you use typical binary classification model and improve the AUC, I do not think it help anything. Must be some out of box thinking that matters.", "476327": "xiaojiu1414\nThank you for the suggestion!\nSo, according to your opinion, outlier detection still works but should be the method different from binary classification? Could l PM you for further suggestion ? (purely intellectual curiosity )", "477261": "Hi @C.Y.W\nSorry for not responding sooner.\nI think the classifier does help,\nOur CV for current LB of 3.668 is about 3.629\n\nWe currently do not use any post processing.\nClassification just adds value to our ensemble.\n\nKeep up the good work!", "478437": "Hi @AmirH,\nI did not use any post processing either in my current solution. And thank you for the reply! \nGood luck to your team!" }, "source": "meta" }