Hi,\nFor me, using the LGBM objective \"poisson\" instead of \"regression\", made a significant improvement.\nAfter testing it, if it works for you, try setting the parameter \"poisson_max_delta_step\" to a higher value, 1.1, 1.5, or even 2, notice that this will require more rounds to train.\n(I think the default value is 0.7)
\n\nLet me know if it improved your model!
", "messages": [ { "id": "343449", "postDate": "06/15/2018 10:41:59", "content": "Hi,\nFor me, using the LGBM objective \"poisson\" instead of \"regression\", made a significant improvement.\nAfter testing it, if it works for you, try setting the parameter \"poisson_max_delta_step\" to a higher value, 1.1, 1.5, or even 2, notice that this will require more rounds to train.\n(I think the default value is 0.7)
\n\nLet me know if it improved your model!
", "rawMarkdown": "Hi,\nFor me, using the LGBM objective \"poisson\" instead of \"regression\", made a significant improvement.\nAfter testing it, if it works for you, try setting the parameter \"poisson_max_delta_step\" to a higher value, 1.1, 1.5, or even 2, notice that this will require more rounds to train.\n(I think the default value is 0.7)\n\nLet me know if it improved your model!", "votes": null }, { "id": "343735", "postDate": "06/16/2018 00:41:19", "content": "No improvement, but terribly destroy my result
", "rawMarkdown": "No improvement, but terribly destroy my result", "votes": null }, { "id": "343775", "postDate": "06/16/2018 03:44:07", "content": "@tpthegreat Can you be more specific.. How much significant improvement.. Did you see an improvement on both Local CV and the LB? I tried a 5-Fold and the local CV is less than what I had earlier by 0.003...
", "rawMarkdown": "tpthegreat Can you be more specific.. How much significant improvement.. Did you see an improvement on both Local CV and the LB? I tried a 5-Fold and the local CV is less than what I had earlier by 0.003...", "votes": null }, { "id": "343818", "postDate": "06/16/2018 07:55:27", "content": "@aquatic @tunguz Did you get any improvement with this approach....
", "rawMarkdown": "aquatic @tunguz Did you get any improvement with this approach....", "votes": null }, { "id": "343883", "postDate": "06/16/2018 12:23:52", "content": "Haven’t tried it yet. I’ll check it out and see if it works.
", "rawMarkdown": "Haven’t tried it yet. I’ll check it out and see if it works.", "votes": null }, { "id": "343946", "postDate": "06/16/2018 15:00:43", "content": "@Samrat @xil lin\nBoth my CV and LB are better by 0.003, without tweeking the poisson delta parameter which gives more improvement.\nI also take the lb predictions and max them down to 1 if they are over, and get them to 0 if they're less.\nI use about 60 features + tfidf columns.\nWeird that it wrecks your results and betters mine. \nBut I'm also much lower ranked than you...
\n\nBtw i also use 5 fold cv
", "rawMarkdown": "Samrat @xil lin\nBoth my CV and LB are better by 0.003, without tweeking the poisson delta parameter which gives more improvement.\nI also take the lb predictions and max them down to 1 if they are over, and get them to 0 if they're less.\nI use about 60 features + tfidf columns.\nWeird that it wrecks your results and betters mine. \nBut I'm also much lower ranked than you... \n\nBtw i also use 5 fold cv", "votes": null }, { "id": "343990", "postDate": "06/16/2018 17:46:20", "content": "It happened to me before on Talking data, after changed from classification to xentropy, it give me extra edges
", "rawMarkdown": "It happened to me before on Talking data, after changed from classification to xentropy, it give me extra edges", "votes": null }, { "id": "345020", "postDate": "06/19/2018 04:39:47", "content": "It dropped my score by 0.0017
", "rawMarkdown": "It dropped my score by 0.0017", "votes": null }, { "id": "345363", "postDate": "06/19/2018 18:39:27", "content": "why so many down-votes for a excellent suggestion ?
", "rawMarkdown": "why so many down-votes for a excellent suggestion ?", "votes": null }, { "id": "345366", "postDate": "06/19/2018 18:44:43", "content": "Yeah, I agree the downvotes are not merited here. This is pretty good advice, even if just for model diversity in ensembling.
", "rawMarkdown": "Yeah, I agree the downvotes are not merited here. This is pretty good advice, even if just for model diversity in ensembling.", "votes": null }, { "id": "345368", "postDate": "06/19/2018 18:49:49", "content": "Yup there's nothing to down vote.. It's a friendly suggestion and if it doesn't work for you that's not his problem..
", "rawMarkdown": "Yup there's nothing to down vote.. It's a friendly suggestion and if it doesn't work for you that's not his problem..", "votes": null }, { "id": "345369", "postDate": "06/19/2018 18:52:15", "content": "Too much negativity on Kaggle these days ...
", "rawMarkdown": "Too much negativity on Kaggle these days ...", "votes": null }, { "id": "345545", "postDate": "06/20/2018 02:50:48", "content": "Can someone experienced tell why changing the objective improve results?
", "rawMarkdown": "Can someone experienced tell why changing the objective improve results?", "votes": null }, { "id": "345786", "postDate": "06/20/2018 13:05:38", "content": "Look at the positive-only targets' distribution, it is a kind of bimodal distribution with two summits of ~0.2 & ~0.8. So , i guess, maybe poisson is a 'better'(or at least different) approximation of the target?
", "rawMarkdown": "Look at the positive-only targets' distribution, it is a kind of bimodal distribution with two summits of ~0.2 & ~0.8. So , i guess, maybe poisson is a 'better'(or at least different) approximation of the target?", "votes": null }, { "id": "347116", "postDate": "06/23/2018 09:30:49", "content": "Thank you for your sharing, but no improvement, even it doubled the time for training.
", "rawMarkdown": "Thank you for your sharing, but no improvement, even it doubled the time for training.", "votes": null } ], "comments": [ { "id": 343735, "author_name": "linendsound", "author_url": "", "post_date": "06/16/2018 00:41:19", "content": "No improvement, but terribly destroy my result
", "votes": null, "replies": [] }, { "id": 343775, "author_name": "samratp", "author_url": "", "post_date": "06/16/2018 03:44:07", "content": "@tpthegreat Can you be more specific.. How much significant improvement.. Did you see an improvement on both Local CV and the LB? I tried a 5-Fold and the local CV is less than what I had earlier by 0.003...
", "votes": null, "replies": [ { "id": 343946, "author_name": "tpthegreat", "author_url": "", "post_date": "06/16/2018 15:00:43", "content": "@Samrat @xil lin\nBoth my CV and LB are better by 0.003, without tweeking the poisson delta parameter which gives more improvement.\nI also take the lb predictions and max them down to 1 if they are over, and get them to 0 if they're less.\nI use about 60 features + tfidf columns.\nWeird that it wrecks your results and betters mine. \nBut I'm also much lower ranked than you...
\n\nBtw i also use 5 fold cv
", "votes": null, "replies": [] } ] }, { "id": 343818, "author_name": "samratp", "author_url": "", "post_date": "06/16/2018 07:55:27", "content": "@aquatic @tunguz Did you get any improvement with this approach....
", "votes": null, "replies": [ { "id": 343883, "author_name": "tunguz", "author_url": "", "post_date": "06/16/2018 12:23:52", "content": "Haven’t tried it yet. I’ll check it out and see if it works.
", "votes": null, "replies": [] } ] }, { "id": 343990, "author_name": "zcistkidd", "author_url": "", "post_date": "06/16/2018 17:46:20", "content": "It happened to me before on Talking data, after changed from classification to xentropy, it give me extra edges
", "votes": null, "replies": [] }, { "id": 345020, "author_name": "nyleve", "author_url": "", "post_date": "06/19/2018 04:39:47", "content": "It dropped my score by 0.0017
", "votes": null, "replies": [] }, { "id": 345363, "author_name": "chabir", "author_url": "", "post_date": "06/19/2018 18:39:27", "content": "why so many down-votes for a excellent suggestion ?
", "votes": null, "replies": [ { "id": 345366, "author_name": "peterhurford", "author_url": "", "post_date": "06/19/2018 18:44:43", "content": "Yeah, I agree the downvotes are not merited here. This is pretty good advice, even if just for model diversity in ensembling.
", "votes": null, "replies": [] }, { "id": 345368, "author_name": "samratp", "author_url": "", "post_date": "06/19/2018 18:49:49", "content": "Yup there's nothing to down vote.. It's a friendly suggestion and if it doesn't work for you that's not his problem..
", "votes": null, "replies": [] }, { "id": 345369, "author_name": "tunguz", "author_url": "", "post_date": "06/19/2018 18:52:15", "content": "Too much negativity on Kaggle these days ...
", "votes": null, "replies": [] } ] }, { "id": 345545, "author_name": "konohayui", "author_url": "", "post_date": "06/20/2018 02:50:48", "content": "Can someone experienced tell why changing the objective improve results?
", "votes": null, "replies": [ { "id": 345786, "author_name": "johnfarrell", "author_url": "", "post_date": "06/20/2018 13:05:38", "content": "Look at the positive-only targets' distribution, it is a kind of bimodal distribution with two summits of ~0.2 & ~0.8. So , i guess, maybe poisson is a 'better'(or at least different) approximation of the target?
", "votes": null, "replies": [] } ] }, { "id": 347116, "author_name": "kownse", "author_url": "", "post_date": "06/23/2018 09:30:49", "content": "Thank you for your sharing, but no improvement, even it doubled the time for training.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "343449": "Hi,\nFor me, using the LGBM objective \"poisson\" instead of \"regression\", made a significant improvement.\nAfter testing it, if it works for you, try setting the parameter \"poisson_max_delta_step\" to a higher value, 1.1, 1.5, or even 2, notice that this will require more rounds to train.\n(I think the default value is 0.7)\n\nLet me know if it improved your model!", "343735": "No improvement, but terribly destroy my result", "343775": "tpthegreat Can you be more specific.. How much significant improvement.. Did you see an improvement on both Local CV and the LB? I tried a 5-Fold and the local CV is less than what I had earlier by 0.003...", "343818": "aquatic @tunguz Did you get any improvement with this approach....", "343883": "Haven’t tried it yet. I’ll check it out and see if it works.", "343946": "Samrat @xil lin\nBoth my CV and LB are better by 0.003, without tweeking the poisson delta parameter which gives more improvement.\nI also take the lb predictions and max them down to 1 if they are over, and get them to 0 if they're less.\nI use about 60 features + tfidf columns.\nWeird that it wrecks your results and betters mine. \nBut I'm also much lower ranked than you... \n\nBtw i also use 5 fold cv", "343990": "It happened to me before on Talking data, after changed from classification to xentropy, it give me extra edges", "345020": "It dropped my score by 0.0017", "345363": "why so many down-votes for a excellent suggestion ?", "345366": "Yeah, I agree the downvotes are not merited here. This is pretty good advice, even if just for model diversity in ensembling.", "345368": "Yup there's nothing to down vote.. It's a friendly suggestion and if it doesn't work for you that's not his problem..", "345369": "Too much negativity on Kaggle these days ...", "345545": "Can someone experienced tell why changing the objective improve results?", "345786": "Look at the positive-only targets' distribution, it is a kind of bimodal distribution with two summits of ~0.2 & ~0.8. So , i guess, maybe poisson is a 'better'(or at least different) approximation of the target?", "347116": "Thank you for your sharing, but no improvement, even it doubled the time for training." }, "source": "meta" }