I'm using LightGBM(LGB) and CatBoost(CAT).
\n\nI realized that the ensemble using my LGB result is somewhat worse than the ensemble using my CAT result.
\n\nDo you have the similar result?
\n\nConsidering the silogram discussion, this phenomenon is caused by the higher over-fitting tendency of LGB.
\n\nBut, that is just my opinion. I want to know it is real and common result. If not, I think there is some mistakes or I need to tune parameters.
", "messages": [ { "id": "399392", "postDate": "10/05/2018 17:43:43", "content": "I'm using LightGBM(LGB) and CatBoost(CAT).
\n\nI realized that the ensemble using my LGB result is somewhat worse than the ensemble using my CAT result.
\n\nDo you have the similar result?
\n\nConsidering the silogram discussion, this phenomenon is caused by the higher over-fitting tendency of LGB.
\n\nBut, that is just my opinion. I want to know it is real and common result. If not, I think there is some mistakes or I need to tune parameters.
", "rawMarkdown": "I'm using LightGBM(LGB) and CatBoost(CAT).\n\nI realized that the ensemble using my LGB result is somewhat worse than the ensemble using my CAT result.\n\nDo you have the similar result?\n\nConsidering the silogram discussion, this phenomenon is caused by the higher over-fitting tendency of LGB.\n\nBut, that is just my opinion. I want to know it is real and common result. If not, I think there is some mistakes or I need to tune parameters.", "votes": null }, { "id": "399396", "postDate": "10/05/2018 17:50:24", "content": "Hi YouHan, would you mind share your CV vs LB score? Have you used the 3 encoding method u share in the kernel you shared before? In my opinion we should compare two models after we finishing tuning hyper parameter...But we know it's time consuming as there is no solid CV right now...We are highly rely on the LB score....
", "rawMarkdown": "Hi YouHan, would you mind share your CV vs LB score? Have you used the 3 encoding method u share in the kernel you shared before? In my opinion we should compare two models after we finishing tuning hyper parameter...But we know it's time consuming as there is no solid CV right now...We are highly rely on the LB score....", "votes": null }, { "id": "399536", "postDate": "10/06/2018 03:45:59", "content": "(1)\nI recognized that my new feature set works when using LGB for now. \nSo, I agree with your comments that I need to tune parameters.
\n\n(2)\nFor now, I'm using frequency and mean encoding because the CV score is lower when using the combination of them.
\n\n(3)
\n\nUnfortunately, I've always used my result as a source of ensemble. So, I don't have LB score using only LGB. Sorry for that!. I need to log all things from now.
\n\nFor now, my LGB CV: 1.5435 LB score (ensemble with my best) : 1.4157
\n\nWithout time features, I got a boost. I recommend you to remove time-related features.
", "rawMarkdown": "(1)\nI recognized that my new feature set works when using LGB for now. \nSo, I agree with your comments that I need to tune parameters.\n\n(2)\nFor now, I'm using frequency and mean encoding because the CV score is lower when using the combination of them.\n\n(3)\n\nUnfortunately, I've always used my result as a source of ensemble. So, I don't have LB score using only LGB. Sorry for that!. I need to log all things from now.\n\nFor now, my LGB CV: 1.5435 LB score (ensemble with my best) : 1.4157\n\nWithout time features, I got a boost. I recommend you to remove time-related features.", "votes": null }, { "id": "399547", "postDate": "10/06/2018 04:22:02", "content": "@youhanlee, have you tuned the CatBoost parameters? Specifically learning_rate & n_estimators.
For now, I didn't tune the parameters. I've focused on feature engineering. \nMaybe catboost will work after parameter tuning.
", "rawMarkdown": "For now, I didn't tune the parameters. I've focused on feature engineering. \nMaybe catboost will work after parameter tuning.", "votes": null }, { "id": "399843", "postDate": "10/06/2018 22:40:30", "content": "Thank you for sharing...I removed some of the time-related features and got a boost... :)
", "rawMarkdown": "Thank you for sharing...I removed some of the time-related features and got a boost... :)", "votes": null } ], "comments": [ { "id": 399396, "author_name": "reputation18", "author_url": "", "post_date": "10/05/2018 17:50:24", "content": "Hi YouHan, would you mind share your CV vs LB score? Have you used the 3 encoding method u share in the kernel you shared before? In my opinion we should compare two models after we finishing tuning hyper parameter...But we know it's time consuming as there is no solid CV right now...We are highly rely on the LB score....
", "votes": null, "replies": [ { "id": 399536, "author_name": "youhanlee", "author_url": "", "post_date": "10/06/2018 03:45:59", "content": "(1)\nI recognized that my new feature set works when using LGB for now. \nSo, I agree with your comments that I need to tune parameters.
\n\n(2)\nFor now, I'm using frequency and mean encoding because the CV score is lower when using the combination of them.
\n\n(3)
\n\nUnfortunately, I've always used my result as a source of ensemble. So, I don't have LB score using only LGB. Sorry for that!. I need to log all things from now.
\n\nFor now, my LGB CV: 1.5435 LB score (ensemble with my best) : 1.4157
\n\nWithout time features, I got a boost. I recommend you to remove time-related features.
", "votes": null, "replies": [] }, { "id": 399547, "author_name": "prashantkikani", "author_url": "", "post_date": "10/06/2018 04:22:02", "content": "@youhanlee, have you tuned the CatBoost parameters? Specifically learning_rate & n_estimators.
For now, I didn't tune the parameters. I've focused on feature engineering. \nMaybe catboost will work after parameter tuning.
", "votes": null, "replies": [] }, { "id": 399843, "author_name": "reputation18", "author_url": "", "post_date": "10/06/2018 22:40:30", "content": "Thank you for sharing...I removed some of the time-related features and got a boost... :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "399392": "I'm using LightGBM(LGB) and CatBoost(CAT).\n\nI realized that the ensemble using my LGB result is somewhat worse than the ensemble using my CAT result.\n\nDo you have the similar result?\n\nConsidering the silogram discussion, this phenomenon is caused by the higher over-fitting tendency of LGB.\n\nBut, that is just my opinion. I want to know it is real and common result. If not, I think there is some mistakes or I need to tune parameters.", "399396": "Hi YouHan, would you mind share your CV vs LB score? Have you used the 3 encoding method u share in the kernel you shared before? In my opinion we should compare two models after we finishing tuning hyper parameter...But we know it's time consuming as there is no solid CV right now...We are highly rely on the LB score....", "399536": "(1)\nI recognized that my new feature set works when using LGB for now. \nSo, I agree with your comments that I need to tune parameters.\n\n(2)\nFor now, I'm using frequency and mean encoding because the CV score is lower when using the combination of them.\n\n(3)\n\nUnfortunately, I've always used my result as a source of ensemble. So, I don't have LB score using only LGB. Sorry for that!. I need to log all things from now.\n\nFor now, my LGB CV: 1.5435 LB score (ensemble with my best) : 1.4157\n\nWithout time features, I got a boost. I recommend you to remove time-related features.", "399547": "youhanlee, have you tuned the CatBoost parameters? Specifically `learning_rate` & `n_estimators`.", "399591": "For now, I didn't tune the parameters. I've focused on feature engineering. \nMaybe catboost will work after parameter tuning.", "399843": "Thank you for sharing...I removed some of the time-related features and got a boost... :)" }, "source": "meta" }