{ "id": 56324, "title": "Dart was a way better in this competition", "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/56324", "author_name": "", "post_date": "2018-05-08T15:31:00.103740400Z", "votes": 2, "comment_count": 12, "views": 0, "content": "

I wonder who else did notice that in this competition choosing a dart for boosting was a much better in term of results? My best submission was made in last day and in was almost basic features set with some counter trained on 1 fold with dar trees

", "messages": [ { "id": "325582", "postDate": "05/08/2018 15:31:00", "content": "

I wonder who else did notice that in this competition choosing a dart for boosting was a much better in term of results? My best submission was made in last day and in was almost basic features set with some counter trained on 1 fold with dar trees

", "rawMarkdown": "I wonder who else did notice that in this competition choosing a dart for boosting was a much better in term of results? My best submission was made in last day and in was almost basic features set with some counter trained on 1 fold with dar trees", "votes": null }, { "id": "325599", "postDate": "05/08/2018 16:05:46", "content": "

we did, our model was mainly based on dart trees - I will do a short write up tonight :)

", "rawMarkdown": "we did, our model was mainly based on dart trees - I will do a short write up tonight :)", "votes": null }, { "id": "325733", "postDate": "05/08/2018 19:49:46", "content": "

Would be great

", "rawMarkdown": "Would be great", "votes": null }, { "id": "325764", "postDate": "05/08/2018 20:46:07", "content": "

I used DART without any other parameter and ended up having bad results. Did you specify how much dropout ? I really thought DART is better than GBDT but could not prove it.

", "rawMarkdown": "I used DART without any other parameter and ended up having bad results. Did you specify how much dropout ? I really thought DART is better than GBDT but could not prove it.", "votes": null }, { "id": "325976", "postDate": "05/09/2018 06:23:58", "content": "

Yes, Dart was better, but slower than the regular one. I used a mix of the two.

", "rawMarkdown": "Yes, Dart was better, but slower than the regular one. I used a mix of the two.", "votes": null }, { "id": "325988", "postDate": "05/09/2018 06:39:38", "content": "

interesting. correct me if I'm wrong, xgboost uses dart as the default boosting method, while lightgbm not. but my lgb outperforms xgb about 0.0004 under the same training data/hyperparameters, would be glad to have some feedback if anyone else tried

", "rawMarkdown": "interesting. correct me if I'm wrong, xgboost uses dart as the default boosting method, while lightgbm not. but my lgb outperforms xgb about 0.0004 under the same training data/hyperparameters, would be glad to have some feedback if anyone else tried", "votes": null }, { "id": "326033", "postDate": "05/09/2018 07:11:35", "content": "

@ValeriyBabushkin, are you asking about booster? http://xgboost.readthedocs.io/en/latest/tutorials/dart.html

", "rawMarkdown": "ValeriyBabushkin, are you asking about booster? http://xgboost.readthedocs.io/en/latest/tutorials/dart.html", "votes": null }, { "id": "326042", "postDate": "05/09/2018 07:23:11", "content": "

what is dart is no hard task to find\nwhy it performs significantly better here is however another question

", "rawMarkdown": "what is dart is no hard task to find\nwhy it performs significantly better here is however another question", "votes": null }, { "id": "326043", "postDate": "05/09/2018 07:25:21", "content": "

Yea, sure. I was just wondering what it is, never heard about it

", "rawMarkdown": "Yea, sure. I was just wondering what it is, never heard about it", "votes": null }, { "id": "326110", "postDate": "05/09/2018 09:55:26", "content": "

I am not sure that default algorithm for xgboost is dart

", "rawMarkdown": "I am not sure that default algorithm for xgboost is dart", "votes": null }, { "id": "326127", "postDate": "05/09/2018 10:21:43", "content": "

booster [default=gbtree] in xgboost

", "rawMarkdown": "booster [default=gbtree] in xgboost", "votes": null }, { "id": "326130", "postDate": "05/09/2018 10:31:25", "content": "

thanks! I got it, would try dart to see if it performs better

", "rawMarkdown": "thanks! I got it, would try dart to see if it performs better", "votes": null }, { "id": "326150", "postDate": "05/09/2018 11:05:05", "content": "

just my experience, dart tree in xgboost before v7 was not very well implemented - lgb's version is much faster and more accurate

", "rawMarkdown": "just my experience, dart tree in xgboost before v7 was not very well implemented - lgb's version is much faster and more accurate", "votes": null } ], "comments": [ { "id": 325599, "author_name": "yifanxie", "author_url": "", "post_date": "05/08/2018 16:05:46", "content": "

we did, our model was mainly based on dart trees - I will do a short write up tonight :)

", "votes": null, "replies": [ { "id": 325733, "author_name": "venheads", "author_url": "", "post_date": "05/08/2018 19:49:46", "content": "

Would be great

", "votes": null, "replies": [] } ] }, { "id": 325764, "author_name": "mayanksoni", "author_url": "", "post_date": "05/08/2018 20:46:07", "content": "

I used DART without any other parameter and ended up having bad results. Did you specify how much dropout ? I really thought DART is better than GBDT but could not prove it.

", "votes": null, "replies": [] }, { "id": 325976, "author_name": "cpmpml", "author_url": "", "post_date": "05/09/2018 06:23:58", "content": "

Yes, Dart was better, but slower than the regular one. I used a mix of the two.

", "votes": null, "replies": [] }, { "id": 325988, "author_name": "atm584", "author_url": "", "post_date": "05/09/2018 06:39:38", "content": "

interesting. correct me if I'm wrong, xgboost uses dart as the default boosting method, while lightgbm not. but my lgb outperforms xgb about 0.0004 under the same training data/hyperparameters, would be glad to have some feedback if anyone else tried

", "votes": null, "replies": [ { "id": 326110, "author_name": "venheads", "author_url": "", "post_date": "05/09/2018 09:55:26", "content": "

I am not sure that default algorithm for xgboost is dart

", "votes": null, "replies": [] }, { "id": 326127, "author_name": "insaff", "author_url": "", "post_date": "05/09/2018 10:21:43", "content": "

booster [default=gbtree] in xgboost

", "votes": null, "replies": [] }, { "id": 326130, "author_name": "atm584", "author_url": "", "post_date": "05/09/2018 10:31:25", "content": "

thanks! I got it, would try dart to see if it performs better

", "votes": null, "replies": [] } ] }, { "id": 326033, "author_name": "insaff", "author_url": "", "post_date": "05/09/2018 07:11:35", "content": "

@ValeriyBabushkin, are you asking about booster? http://xgboost.readthedocs.io/en/latest/tutorials/dart.html

", "votes": null, "replies": [ { "id": 326042, "author_name": "venheads", "author_url": "", "post_date": "05/09/2018 07:23:11", "content": "

what is dart is no hard task to find\nwhy it performs significantly better here is however another question

", "votes": null, "replies": [] }, { "id": 326043, "author_name": "insaff", "author_url": "", "post_date": "05/09/2018 07:25:21", "content": "

Yea, sure. I was just wondering what it is, never heard about it

", "votes": null, "replies": [] } ] }, { "id": 326150, "author_name": "yifanxie", "author_url": "", "post_date": "05/09/2018 11:05:05", "content": "

just my experience, dart tree in xgboost before v7 was not very well implemented - lgb's version is much faster and more accurate

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "325582": "I wonder who else did notice that in this competition choosing a dart for boosting was a much better in term of results? My best submission was made in last day and in was almost basic features set with some counter trained on 1 fold with dar trees", "325599": "we did, our model was mainly based on dart trees - I will do a short write up tonight :)", "325733": "Would be great", "325764": "I used DART without any other parameter and ended up having bad results. Did you specify how much dropout ? I really thought DART is better than GBDT but could not prove it.", "325976": "Yes, Dart was better, but slower than the regular one. I used a mix of the two.", "325988": "interesting. correct me if I'm wrong, xgboost uses dart as the default boosting method, while lightgbm not. but my lgb outperforms xgb about 0.0004 under the same training data/hyperparameters, would be glad to have some feedback if anyone else tried", "326033": "ValeriyBabushkin, are you asking about booster? http://xgboost.readthedocs.io/en/latest/tutorials/dart.html", "326042": "what is dart is no hard task to find\nwhy it performs significantly better here is however another question", "326043": "Yea, sure. I was just wondering what it is, never heard about it", "326110": "I am not sure that default algorithm for xgboost is dart", "326127": "booster [default=gbtree] in xgboost", "326130": "thanks! I got it, would try dart to see if it performs better", "326150": "just my experience, dart tree in xgboost before v7 was not very well implemented - lgb's version is much faster and more accurate" }, "source": "meta" }