![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5419716%2F7297d2b5ce2bc14eb3b6fb783feab3fc%2FCapturar.PNG?generation=1683853771340534&alt=media\")
Using 300 features gave CV:0.698, but LB was 0.646. Using 33 features gave CV:0.679, and LB was 0.679. I used Chris' xgboost Baseline.
", "messages": [ { "id": 2148389, "postDate": "2023-02-17T12:04:52.417Z", "content": "Using 300 features gave CV:0.698, but LB was 0.646. Using 33 features gave CV:0.679, and LB was 0.679. I used Chris' xgboost Baseline.
", "rawMarkdown": "Using 300 features gave CV:0.698, but LB was 0.646. Using 33 features gave CV:0.679, and LB was 0.679. I used Chris' xgboost Baseline. ", "votes": 11 }, { "id": 2148413, "postDate": "2023-02-17T12:28:58.357Z", "content": "You may check the features importance of the model. There may be a feature/s that introduced a leakage. If you found the model depends too much on one of the features, try to drop it.
", "rawMarkdown": "You may check the features importance of the model. There may be a feature/s that introduced a leakage. If you found the model depends too much on one of the features, try to drop it.", "votes": 3, "replies": [ { "id": 2148431, "postDate": "2023-02-17T12:47:19.693Z", "rawMarkdown": "", "votes": 1, "isDeleted": true, "replies": [ { "id": 2148449, "postDate": "2023-02-17T12:57:52.553Z", "content": "I think it is not a problem as long as it is possible to be done on the test set and the target is not involved. Btw, checking the features importance is best thing you can do. If you found many features have high importance, try to drop each of the top ones separately and rerun the model. You may then discover what caused the problem. Sometimes there is one feature that decreases the performance significantly. Not sure, but this usually happens for me with some of the lags and diffs features when I add them recklessly.
", "rawMarkdown": "I think it is not a problem as long as it is possible to be done on the test set and the target is not involved. Btw, checking the features importance is best thing you can do. If you found many features have high importance, try to drop each of the top ones separately and rerun the model. You may then discover what caused the problem. Sometimes there is one feature that decreases the performance significantly. Not sure, but this usually happens for me with some of the lags and diffs features when I add them recklessly.", "votes": 1 }, { "id": 2148502, "postDate": "2023-02-17T13:43:26.573Z", "content": "I observed that the feature importance of this feature is about twice as high as other features for several questions.
", "rawMarkdown": "I observed that the feature importance of this feature is about twice as high as other features for several questions. " }, { "id": 2148545, "postDate": "2023-02-17T14:23:47.453Z", "content": "Try to drop it and notice if the cv and lb correlate again.
", "rawMarkdown": "Try to drop it and notice if the cv and lb correlate again." }, { "id": 2149383, "postDate": "2023-02-18T08:35:58.870Z", "content": "It appears that I used diff.shift(-1), so the data for level_group 5-12 was leaked when calculating the features for level_group 1-4. After fixing this bug, the CV now correlates to LB.
", "rawMarkdown": "It appears that I used diff.shift(-1), so the data for level_group 5-12 was leaked when calculating the features for level_group 1-4. After fixing this bug, the CV now correlates to LB.", "votes": 3 } ] }, { "id": 2255750, "postDate": "2023-05-12T01:11:36.180Z", "content": "
@mohammad2012191 I'm currently having a gap of 0.003, should I drop these columns with a high score?
", "rawMarkdown": "\n\n@mohammad2012191 I'm currently having a gap of 0.003, should I drop these columns with a high score?" } ] }, { "id": 2149140, "postDate": "2023-02-18T01:49:38.690Z", "content": "For me, big boost in cv brings big boost in LB. But small boost in cv sometimes didn't work for LB.
", "rawMarkdown": "For me, big boost in cv brings big boost in LB. But small boost in cv sometimes didn't work for LB.", "votes": 1 }, { "id": 2148495, "postDate": "2023-02-17T13:32:47.897Z", "content": "For me CV doesnt correlate perfectly with LB but pretty close.
", "rawMarkdown": "For me CV doesnt correlate perfectly with LB but pretty close. " } ], "comments": [ { "id": 2148413, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2023-02-17T12:28:58.357000", "content": "You may check the features importance of the model. There may be a feature/s that introduced a leakage. If you found the model depends too much on one of the features, try to drop it.
", "votes": 3, "replies": [ { "id": 2148431, "author_name": "", "author_url": "", "post_date": "2023-02-17T12:47:19.693000", "content": "", "votes": 1, "replies": [ { "id": 2148449, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2023-02-17T12:57:52.553000", "content": "I think it is not a problem as long as it is possible to be done on the test set and the target is not involved. Btw, checking the features importance is best thing you can do. If you found many features have high importance, try to drop each of the top ones separately and rerun the model. You may then discover what caused the problem. Sometimes there is one feature that decreases the performance significantly. Not sure, but this usually happens for me with some of the lags and diffs features when I add them recklessly.
", "votes": 1, "replies": [] }, { "id": 2148502, "author_name": "theta", "author_url": "", "post_date": "2023-02-17T13:43:26.573000", "content": "I observed that the feature importance of this feature is about twice as high as other features for several questions.
", "votes": 0, "replies": [] }, { "id": 2148545, "author_name": "Mohamed Eltayeb", "author_url": "", "post_date": "2023-02-17T14:23:47.453000", "content": "Try to drop it and notice if the cv and lb correlate again.
", "votes": 0, "replies": [] }, { "id": 2149383, "author_name": "theta", "author_url": "", "post_date": "2023-02-18T08:35:58.870000", "content": "It appears that I used diff.shift(-1), so the data for level_group 5-12 was leaked when calculating the features for level_group 1-4. After fixing this bug, the CV now correlates to LB.
", "votes": 3, "replies": [] } ] }, { "id": 2255750, "author_name": "adriel cabral", "author_url": "", "post_date": "2023-05-12T01:11:36.180000", "content": "
@mohammad2012191 I'm currently having a gap of 0.003, should I drop these columns with a high score?
", "votes": 0, "replies": [] } ] }, { "id": 2149140, "author_name": "ADAM.", "author_url": "", "post_date": "2023-02-18T01:49:38.690000", "content": "For me, big boost in cv brings big boost in LB. But small boost in cv sometimes didn't work for LB.
", "votes": 1, "replies": [] }, { "id": 2148495, "author_name": "Simon Veitner", "author_url": "", "post_date": "2023-02-17T13:32:47.897000", "content": "For me CV doesnt correlate perfectly with LB but pretty close.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2148389": "Using 300 features gave CV:0.698, but LB was 0.646. Using 33 features gave CV:0.679, and LB was 0.679. I used Chris' xgboost Baseline. ", "2148413": "You may check the features importance of the model. There may be a feature/s that introduced a leakage. If you found the model depends too much on one of the features, try to drop it.", "2149140": "For me, big boost in cv brings big boost in LB. But small boost in cv sometimes didn't work for LB.", "2148495": "For me CV doesnt correlate perfectly with LB but pretty close. " } }