![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4516327%2F85f599536307412872f18989105ee547%2Fnewplot%20(2).png?generation=1710510274233825&alt=media\")
I noticed that some features without the D suffix represent the month and year, so their distribution changes depending on time. Since decision tree-based models cannot perform extrapolation, such features harm model performance. You can find them by searching for \"year\" and \"month\" in the column names.
\n
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4516327%2F507070a839a91c21297dc74d52c81bb5%2Fnewplot%20(1).png?generation=1710510526060588&alt=media\")
I noticed that some features without the D suffix represent the month and year, so their distribution changes depending on time. Since decision tree-based models cannot perform extrapolation, such features harm model performance. You can find them by searching for \"year\" and \"month\" in the column names.
\n
How should we deal with such features?
", "rawMarkdown": "How should we deal with such features?", "votes": null }, { "id": "2704664", "postDate": "03/18/2024 21:00:52", "content": "You can subtract from prediction date (date_decision). In such cases, you can prevent feature drift by setting the date of the prediction as base point
", "rawMarkdown": "You can subtract from prediction date (date_decision). In such cases, you can prevent feature drift by setting the date of the prediction as base point", "votes": null }, { "id": "2715418", "postDate": "03/25/2024 13:34:28", "content": "Does the \"Test dataset\" in first figure refer to public test table?
", "rawMarkdown": "Does the \"Test dataset\" in first figure refer to public test table?", "votes": null }, { "id": "2715977", "postDate": "03/25/2024 18:58:58", "content": "I configured weeks 0-45 of the public data as train, and weeks 45-90 as test data set
", "rawMarkdown": "I configured weeks 0-45 of the public data as train, and weeks 45-90 as test data set", "votes": null }, { "id": "2716028", "postDate": "03/25/2024 19:27:48", "content": "Excuse me for interrupting. Did you find any correlation between Local Score and LB when implementing the Time Series data split method?
\nI added feature engineering to the provided notebook (using 5-fold cross-validation without considering time series), but CV and LB did not correlate.
", "rawMarkdown": "Excuse me for interrupting. Did you find any correlation between Local Score and LB when implementing the Time Series data split method?\n\nI added feature engineering to the provided notebook (using 5-fold cross-validation without considering time series), but CV and LB did not correlate.", "votes": null }, { "id": "2717113", "postDate": "03/26/2024 11:31:20", "content": "Unfortunately I couldn't find any correlation. Now I'm starting from scratch. Scores over 0.540 can be achieved with static tables
", "rawMarkdown": "Unfortunately I couldn't find any correlation. Now I'm starting from scratch. Scores over 0.540 can be achieved with static tables", "votes": null } ], "comments": [ { "id": 2703826, "author_name": "uiosun", "author_url": "", "post_date": "03/18/2024 12:30:43", "content": "How should we deal with such features?
", "votes": null, "replies": [ { "id": 2704664, "author_name": "greysky", "author_url": "", "post_date": "03/18/2024 21:00:52", "content": "You can subtract from prediction date (date_decision). In such cases, you can prevent feature drift by setting the date of the prediction as base point
", "votes": null, "replies": [] } ] }, { "id": 2715418, "author_name": "dimakoshman", "author_url": "", "post_date": "03/25/2024 13:34:28", "content": "Does the \"Test dataset\" in first figure refer to public test table?
", "votes": null, "replies": [ { "id": 2715977, "author_name": "greysky", "author_url": "", "post_date": "03/25/2024 18:58:58", "content": "I configured weeks 0-45 of the public data as train, and weeks 45-90 as test data set
", "votes": null, "replies": [ { "id": 2716028, "author_name": "youheitomio", "author_url": "", "post_date": "03/25/2024 19:27:48", "content": "Excuse me for interrupting. Did you find any correlation between Local Score and LB when implementing the Time Series data split method?
\nI added feature engineering to the provided notebook (using 5-fold cross-validation without considering time series), but CV and LB did not correlate.
", "votes": null, "replies": [ { "id": 2717113, "author_name": "greysky", "author_url": "", "post_date": "03/26/2024 11:31:20", "content": "Unfortunately I couldn't find any correlation. Now I'm starting from scratch. Scores over 0.540 can be achieved with static tables
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "2698418": "I noticed that some features without the D suffix represent the month and year, so their distribution changes depending on time. Since decision tree-based models cannot perform extrapolation, such features harm model performance. You can find them by searching for \"year\" and \"month\" in the column names.\n\n.png?generation=1710510274233825&alt=media)\n\n.png?generation=1710510526060588&alt=media)", "2703826": "How should we deal with such features?", "2704664": "You can subtract from prediction date (date_decision). In such cases, you can prevent feature drift by setting the date of the prediction as base point", "2715418": "Does the \"Test dataset\" in first figure refer to public test table?", "2715977": "I configured weeks 0-45 of the public data as train, and weeks 45-90 as test data set", "2716028": "Excuse me for interrupting. Did you find any correlation between Local Score and LB when implementing the Time Series data split method?\n\nI added feature engineering to the provided notebook (using 5-fold cross-validation without considering time series), but CV and LB did not correlate.", "2717113": "Unfortunately I couldn't find any correlation. Now I'm starting from scratch. Scores over 0.540 can be achieved with static tables" }, "source": "meta" }