Why are these users are considered as 'is_churn=1' in March?
\n(Please check attached links - churn labels and transaction histories for three users)
Last two examples are what previous Kaggler(@yliu) has been reported.
\n\nI have to predict the prediction label, but I can not trust labels in the train data yet.\nThe label of the unveiled test data also seems to have the same problem.
\n\n(NOTE: A 'transactions' dataframe is a concatenation of two transactions data given by the competition)
", "messages": [ { "id": 250080, "postDate": "2017-11-29T18:26:52.243Z", "content": "I think you can reduce the difference between the extracted files and the official ones if you change the variable historyCutoff to some months in the past according to Arden Chiu in this post.
\n\nEven doing that I was not able to exactly match the official train files...
", "rawMarkdown": "I think you can reduce the difference between the extracted files and the official ones if you change the variable **historyCutoff** to some months in the past according to Arden Chiu in [this post][1].\n\nEven doing that I was not able to exactly match the official train files... \n\n [1]: https://www.kaggle.com/c/kkbox-churn-prediction-challenge/discussion/43145", "votes": 1 }, { "id": 249983, "postDate": "2017-11-29T15:38:55.433Z", "content": "The first one you mentioned (Msno=hDW0/cSgIayezNOh5RsbKKskn4WsNUhrx6ZEbInTSQk=) is definitely a mistake of some kind. According to the scala labeller code, it should not have been included at all in the march test set (train2).
\n\nScala Explanation: Using March date cutoffs (historyCutoff=20170228), this msno in the \"historyData\" dataset would have had a \"last_expire\" date of 20170409, which means it is filtered out of the \"predictionCandidates\" dataset which only includes msno with a last_expire date between 20170301 and after 20170331.
\n\nClearly the scala labeller was either ran with incorrect values (improper date cutoffs) to generate our train/test sets, or the code we were given was not the exact code actually used.
", "rawMarkdown": "The first one you mentioned (Msno=hDW0/cSgIayezNOh5RsbKKskn4WsNUhrx6ZEbInTSQk=) is definitely a mistake of some kind. According to the scala labeller code, it should not have been included at all in the march test set (train2). \n\nScala Explanation: Using March date cutoffs (historyCutoff=20170228), this msno in the \"historyData\" dataset would have had a \"last_expire\" date of 20170409, which means it is filtered out of the \"predictionCandidates\" dataset which only includes msno with a last_expire date between 20170301 and after 20170331.\n\nClearly the scala labeller was either ran with incorrect values (improper date cutoffs) to generate our train/test sets, or the code we were given was not the exact code actually used.", "votes": 1, "replies": [ { "id": 250810, "postDate": "2017-11-30T10:52:24.607Z", "content": "This is really helpful for me to understand the Scala labeler which I was not familiar with. \nI wasn't considering the labeler yet, but the key aspect to get the high score seems like doing optimization according to labels from this labeler...!
", "rawMarkdown": "This is really helpful for me to understand the Scala labeler which I was not familiar with. \nI wasn't considering the labeler yet, but the key aspect to get the high score seems like doing optimization according to labels from this labeler...! " } ] }, { "id": 248549, "postDate": "2017-11-26T13:05:44.653Z", "content": "Why are these users are considered as 'is_churn=1' in March?
\n(Please check attached links - churn labels and transaction histories for three users)
Last two examples are what previous Kaggler(@yliu) has been reported.
\n\nI have to predict the prediction label, but I can not trust labels in the train data yet.\nThe label of the unveiled test data also seems to have the same problem.
\n\n(NOTE: A 'transactions' dataframe is a concatenation of two transactions data given by the competition)
", "rawMarkdown": "Why are these users are considered as 'is_churn=1' in March?I think you can reduce the difference between the extracted files and the official ones if you change the variable historyCutoff to some months in the past according to Arden Chiu in this post.
\n\nEven doing that I was not able to exactly match the official train files...
", "votes": 1, "replies": [] }, { "id": 249983, "author_name": "Bryan Gregory", "author_url": "", "post_date": "2017-11-29T15:38:55.433000", "content": "The first one you mentioned (Msno=hDW0/cSgIayezNOh5RsbKKskn4WsNUhrx6ZEbInTSQk=) is definitely a mistake of some kind. According to the scala labeller code, it should not have been included at all in the march test set (train2).
\n\nScala Explanation: Using March date cutoffs (historyCutoff=20170228), this msno in the \"historyData\" dataset would have had a \"last_expire\" date of 20170409, which means it is filtered out of the \"predictionCandidates\" dataset which only includes msno with a last_expire date between 20170301 and after 20170331.
\n\nClearly the scala labeller was either ran with incorrect values (improper date cutoffs) to generate our train/test sets, or the code we were given was not the exact code actually used.
", "votes": 1, "replies": [ { "id": 250810, "author_name": "Sundong Kim", "author_url": "", "post_date": "2017-11-30T10:52:24.607000", "content": "This is really helpful for me to understand the Scala labeler which I was not familiar with. \nI wasn't considering the labeler yet, but the key aspect to get the high score seems like doing optimization according to labels from this labeler...!
", "votes": 0, "replies": [] } ] }, { "id": 249961, "author_name": "", "author_url": "", "post_date": "2017-11-29T15:02:35.903000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "250080": "I think you can reduce the difference between the extracted files and the official ones if you change the variable **historyCutoff** to some months in the past according to Arden Chiu in [this post][1].\n\nEven doing that I was not able to exactly match the official train files... \n\n [1]: https://www.kaggle.com/c/kkbox-churn-prediction-challenge/discussion/43145", "249983": "The first one you mentioned (Msno=hDW0/cSgIayezNOh5RsbKKskn4WsNUhrx6ZEbInTSQk=) is definitely a mistake of some kind. According to the scala labeller code, it should not have been included at all in the march test set (train2). \n\nScala Explanation: Using March date cutoffs (historyCutoff=20170228), this msno in the \"historyData\" dataset would have had a \"last_expire\" date of 20170409, which means it is filtered out of the \"predictionCandidates\" dataset which only includes msno with a last_expire date between 20170301 and after 20170331.\n\nClearly the scala labeller was either ran with incorrect values (improper date cutoffs) to generate our train/test sets, or the code we were given was not the exact code actually used.", "248549": "Why are these users are considered as 'is_churn=1' in March?