{ "id": 80025, "title": "Is score below LB 3.675 feasible with single LGB?", "url": "/competitions/elo-merchant-category-recommendation/discussion/80025", "author_name": "SubikashPal", "post_date": "2019-02-09T16:24:16.834000", "votes": 4, "comment_count": 28, "views": 0, "content": "

I am struggling to improve my single model beyond 3.677. Any body who crossed the limit?

", "messages": [ { "id": 468793, "postDate": "2019-02-09T17:17:43.250Z", "content": "

my best model is at 3.688. just increase the diversity of your models :)

", "rawMarkdown": "my best model is at 3.688. just increase the diversity of your models :)", "votes": 8, "replies": [ { "id": 468797, "postDate": "2019-02-09T17:29:47.877Z", "content": "

May I know what diversity means? Sorry I am a newbie to the competition, and my exposure is limited. I hope you don't mind enlightening me with your hard-earned experience.

", "rawMarkdown": "May I know what diversity means? Sorry I am a newbie to the competition, and my exposure is limited. I hope you don't mind enlightening me with your hard-earned experience. " }, { "id": 468800, "postDate": "2019-02-09T17:36:25.030Z", "content": "

Its building many different models on different feature sets, then taking the prediction average (or weighted average). This is called ensembling

\n\n

Having many weak models ensembled sometimes may be better than having a few very strong models

", "rawMarkdown": "Its building many different models on different feature sets, then taking the prediction average (or weighted average). This is called ensembling\n\nHaving many weak models ensembled sometimes may be better than having a few very strong models", "votes": 5 }, { "id": 468805, "postDate": "2019-02-09T18:02:27.573Z", "content": "

what method did you use to combine them and what types of models are you using?\nAlso, how many feature sets did you use, and how many different models have you combined.

", "rawMarkdown": "what method did you use to combine them and what types of models are you using?\nAlso, how many feature sets did you use, and how many different models have you combined.\n" }, { "id": 468867, "postDate": "2019-02-09T22:30:00.663Z", "content": "

@radar , Sir, how may diverse model you are having currently ?

", "rawMarkdown": "@radar , Sir, how may diverse model you are having currently ?", "votes": 1 }, { "id": 468871, "postDate": "2019-02-09T22:36:04.090Z", "content": "

curious to know too

", "rawMarkdown": "curious to know too" }, { "id": 468874, "postDate": "2019-02-09T22:57:52.553Z", "content": "

my best single LGBM is at LB3.670 (CV3.622, 6fold, 230features, just feature engineering). Ensemble yet.
\nI feel I saw Yifan has 3.668 in the discussion.
\nAs raddar says, we should increase the diversity of the model. I try it.

", "rawMarkdown": "my best single LGBM is at LB3.670 (CV3.622, 6fold, 230features, just feature engineering). Ensemble yet. \nI feel I saw Yifan has 3.668 in the discussion. \nAs raddar says, we should increase the diversity of the model. I try it. ", "votes": 1 }, { "id": 468923, "postDate": "2019-02-10T03:32:38.883Z", "content": "

3.670? wahoooo.. excellent.\ncouple of questions.\ndid u use merchant.csv?\nare these only aggregated features? or used any other sort of algo.\nany advice..

", "rawMarkdown": "3.670? wahoooo.. excellent.\ncouple of questions.\ndid u use merchant.csv?\nare these only aggregated features? or used any other sort of algo.\nany advice.." }, { "id": 468926, "postDate": "2019-02-10T03:33:21.270Z", "content": "

Congratulations radar.awesome different models for sure.

", "rawMarkdown": "Congratulations radar.awesome different models for sure." }, { "id": 468988, "postDate": "2019-02-10T08:13:06.277Z", "content": "

did u use merchant.csv?
\n-> No.\nare these only aggregated features? or used any other sort of algo.
\n-> I made many various aggregated features. no algo. Feature selection is done carefully. I think about historical behavior features. I don't know about overfit.

", "rawMarkdown": "did u use merchant.csv? \n-> No.\nare these only aggregated features? or used any other sort of algo. \n-> I made many various aggregated features. no algo. Feature selection is done carefully. I think about historical behavior features. I don't know about overfit. " }, { "id": 468992, "postDate": "2019-02-10T08:42:57.800Z", "content": "

are u planning to team up?

", "rawMarkdown": "are u planning to team up?" }, { "id": 468998, "postDate": "2019-02-10T09:20:37.800Z", "content": "

I am still considering it. but, at least, it is difficult with you by submission limit.

\n\n
\n\n

Beyond that I mistook the place to reply...sorry, @raddar
\nI follow your advice, I make a variety of models. I am looking forward to hearing what model you are training.

", "rawMarkdown": "I am still considering it. but, at least, it is difficult with you by submission limit. \n- - - - - \nBeyond that I mistook the place to reply...sorry, @raddar \nI follow your advice, I make a variety of models. I am looking forward to hearing what model you are training." }, { "id": 469026, "postDate": "2019-02-10T10:42:33.937Z", "content": "

i still don't have success with ensembling. all my other models perform very poorly as compared to the single one that gave my 3.686 LB. The others are 3.7+

", "rawMarkdown": "i still don't have success with ensembling. all my other models perform very poorly as compared to the single one that gave my 3.686 LB. The others are 3.7+" }, { "id": 474225, "postDate": "2019-02-19T05:11:21.677Z", "content": "

3.668 is break by my team now ,we got 3.667 best sigle model

", "rawMarkdown": "3.668 is break by my team now ,we got 3.667 best sigle model", "votes": 3 }, { "id": 474228, "postDate": "2019-02-19T05:16:28.717Z", "content": "

truly cool! may I ask that if all this improvement made by feature engineering or parameter tuning?

", "rawMarkdown": "truly cool! may I ask that if all this improvement made by feature engineering or parameter tuning?" }, { "id": 474308, "postDate": "2019-02-19T07:34:50.643Z", "content": "

sigle model? what's that? i think you should spend more time on FE

", "rawMarkdown": "sigle model? what's that? i think you should spend more time on FE" }, { "id": 474492, "postDate": "2019-02-19T12:38:24.553Z", "content": "

wahoo...how are u getting so much of features?

", "rawMarkdown": "wahoo...how are u getting so much of features?" } ] }, { "id": 468776, "postDate": "2019-02-09T16:24:16.833Z", "content": "

I am struggling to improve my single model beyond 3.677. Any body who crossed the limit?

", "rawMarkdown": "I am struggling to improve my single model beyond 3.677. Any body who crossed the limit?", "votes": 4 }, { "id": 472966, "postDate": "2019-02-17T02:09:54.883Z", "content": "

wondering now

", "rawMarkdown": "wondering now" }, { "id": 472832, "postDate": "2019-02-16T18:14:50.433Z", "content": "

I see you @SubikashPal. So you're starting to do some serious outlier prediction now huh?

", "rawMarkdown": "I see you @SubikashPal. So you're starting to do some serious outlier prediction now huh? ", "replies": [ { "id": 472853, "postDate": "2019-02-16T19:10:31.090Z", "content": "

Yah..there's nothing i can do in terms of single model. I am mainly working on stacking and outlier.\nbtw - your score is quite encouraging? Are u able to improve your single model?

", "rawMarkdown": "Yah..there's nothing i can do in terms of single model. I am mainly working on stacking and outlier.\nbtw - your score is quite encouraging? Are u able to improve your single model?" }, { "id": 472875, "postDate": "2019-02-16T20:07:52.860Z", "content": "

I'm still stuck at 3.684 LB for my best model, but my CV's been improving a little. It may have improved, at this point I'm not wasting submissions on a single model anymore.

", "rawMarkdown": "I'm still stuck at 3.684 LB for my best model, but my CV's been improving a little. It may have improved, at this point I'm not wasting submissions on a single model anymore." } ] }, { "id": 472756, "postDate": "2019-02-16T16:02:32.570Z", "content": "

I just got the 3.679 and wondering how could someone reaches to 3.670???

", "rawMarkdown": "I just got the 3.679 and wondering how could someone reaches to 3.670???", "replies": [ { "id": 472827, "postDate": "2019-02-16T17:59:28.533Z", "content": "

even i wonder.

", "rawMarkdown": "even i wonder." } ] }, { "id": 468908, "postDate": "2019-02-10T02:26:00.977Z", "content": "

My best single model is CV 3.663 and LB 3.692 . Does anyone give me advice?

", "rawMarkdown": "My best single model is CV 3.663 and LB 3.692 . Does anyone give me advice?" }, { "id": 468858, "postDate": "2019-02-09T21:35:10.957Z", "content": "

So far i have come across best single model scoring LB 3.674 reported by @LongYin \nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73517#462946

", "rawMarkdown": "So far i have come across best single model scoring LB 3.674 reported by @LongYin \nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73517#462946" }, { "id": 468780, "postDate": "2019-02-09T16:40:03.697Z", "content": "

I can't even get past 3.687. So frustrating.

", "rawMarkdown": "I can't even get past 3.687. So frustrating.", "replies": [ { "id": 468781, "postDate": "2019-02-09T16:43:56.103Z", "content": "

use some sorts of aggregation based on Category. Should give some solid features.

", "rawMarkdown": "use some sorts of aggregation based on Category. Should give some solid features.", "votes": 2 } ] }, { "id": 468900, "postDate": "2019-02-10T01:37:48.003Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 468793, "author_name": "raddar", "author_url": "", "post_date": "2019-02-09T17:17:43.250000", "content": "

my best model is at 3.688. just increase the diversity of your models :)

", "votes": 8, "replies": [ { "id": 468797, "author_name": "FP", "author_url": "", "post_date": "2019-02-09T17:29:47.877000", "content": "

May I know what diversity means? Sorry I am a newbie to the competition, and my exposure is limited. I hope you don't mind enlightening me with your hard-earned experience.

", "votes": 0, "replies": [] }, { "id": 468800, "author_name": "raddar", "author_url": "", "post_date": "2019-02-09T17:36:25.030000", "content": "

Its building many different models on different feature sets, then taking the prediction average (or weighted average). This is called ensembling

\n\n

Having many weak models ensembled sometimes may be better than having a few very strong models

", "votes": 5, "replies": [] }, { "id": 468805, "author_name": "Kyle Peters", "author_url": "", "post_date": "2019-02-09T18:02:27.573000", "content": "

what method did you use to combine them and what types of models are you using?\nAlso, how many feature sets did you use, and how many different models have you combined.

", "votes": 0, "replies": [] }, { "id": 468867, "author_name": "Manoj", "author_url": "", "post_date": "2019-02-09T22:30:00.663000", "content": "

@radar , Sir, how may diverse model you are having currently ?

", "votes": 1, "replies": [] }, { "id": 468871, "author_name": "FP", "author_url": "", "post_date": "2019-02-09T22:36:04.090000", "content": "

curious to know too

", "votes": 0, "replies": [] }, { "id": 468874, "author_name": "holygo_1208", "author_url": "", "post_date": "2019-02-09T22:57:52.553000", "content": "

my best single LGBM is at LB3.670 (CV3.622, 6fold, 230features, just feature engineering). Ensemble yet.
\nI feel I saw Yifan has 3.668 in the discussion.
\nAs raddar says, we should increase the diversity of the model. I try it.

", "votes": 1, "replies": [] }, { "id": 468923, "author_name": "SubikashPal", "author_url": "", "post_date": "2019-02-10T03:32:38.883000", "content": "

3.670? wahoooo.. excellent.\ncouple of questions.\ndid u use merchant.csv?\nare these only aggregated features? or used any other sort of algo.\nany advice..

", "votes": 0, "replies": [] }, { "id": 468926, "author_name": "SubikashPal", "author_url": "", "post_date": "2019-02-10T03:33:21.270000", "content": "

Congratulations radar.awesome different models for sure.

", "votes": 0, "replies": [] }, { "id": 468988, "author_name": "holygo_1208", "author_url": "", "post_date": "2019-02-10T08:13:06.277000", "content": "

did u use merchant.csv?
\n-> No.\nare these only aggregated features? or used any other sort of algo.
\n-> I made many various aggregated features. no algo. Feature selection is done carefully. I think about historical behavior features. I don't know about overfit.

", "votes": 0, "replies": [] }, { "id": 468992, "author_name": "SubikashPal", "author_url": "", "post_date": "2019-02-10T08:42:57.800000", "content": "

are u planning to team up?

", "votes": 0, "replies": [] }, { "id": 468998, "author_name": "holygo_1208", "author_url": "", "post_date": "2019-02-10T09:20:37.800000", "content": "

I am still considering it. but, at least, it is difficult with you by submission limit.

\n\n
\n\n

Beyond that I mistook the place to reply...sorry, @raddar
\nI follow your advice, I make a variety of models. I am looking forward to hearing what model you are training.

", "votes": 0, "replies": [] }, { "id": 469026, "author_name": "dylonLL", "author_url": "", "post_date": "2019-02-10T10:42:33.937000", "content": "

i still don't have success with ensembling. all my other models perform very poorly as compared to the single one that gave my 3.686 LB. The others are 3.7+

", "votes": 0, "replies": [] }, { "id": 474225, "author_name": "Orange", "author_url": "", "post_date": "2019-02-19T05:11:21.677000", "content": "

3.668 is break by my team now ,we got 3.667 best sigle model

", "votes": 3, "replies": [] }, { "id": 474228, "author_name": "Payne zhu", "author_url": "", "post_date": "2019-02-19T05:16:28.717000", "content": "

truly cool! may I ask that if all this improvement made by feature engineering or parameter tuning?

", "votes": 0, "replies": [] }, { "id": 474308, "author_name": "process531", "author_url": "", "post_date": "2019-02-19T07:34:50.643000", "content": "

sigle model? what's that? i think you should spend more time on FE

", "votes": 0, "replies": [] }, { "id": 474492, "author_name": "SubikashPal", "author_url": "", "post_date": "2019-02-19T12:38:24.553000", "content": "

wahoo...how are u getting so much of features?

", "votes": 0, "replies": [] } ] }, { "id": 472966, "author_name": "Payne zhu", "author_url": "", "post_date": "2019-02-17T02:09:54.883000", "content": "

wondering now

", "votes": 0, "replies": [] }, { "id": 472832, "author_name": "Mendrika Ramarlina", "author_url": "", "post_date": "2019-02-16T18:14:50.433000", "content": "

I see you @SubikashPal. So you're starting to do some serious outlier prediction now huh?

", "votes": 0, "replies": [ { "id": 472853, "author_name": "SubikashPal", "author_url": "", "post_date": "2019-02-16T19:10:31.090000", "content": "

Yah..there's nothing i can do in terms of single model. I am mainly working on stacking and outlier.\nbtw - your score is quite encouraging? Are u able to improve your single model?

", "votes": 0, "replies": [] }, { "id": 472875, "author_name": "Mendrika Ramarlina", "author_url": "", "post_date": "2019-02-16T20:07:52.860000", "content": "

I'm still stuck at 3.684 LB for my best model, but my CV's been improving a little. It may have improved, at this point I'm not wasting submissions on a single model anymore.

", "votes": 0, "replies": [] } ] }, { "id": 472756, "author_name": "Link", "author_url": "", "post_date": "2019-02-16T16:02:32.570000", "content": "

I just got the 3.679 and wondering how could someone reaches to 3.670???

", "votes": 0, "replies": [ { "id": 472827, "author_name": "SubikashPal", "author_url": "", "post_date": "2019-02-16T17:59:28.533000", "content": "

even i wonder.

", "votes": 0, "replies": [] } ] }, { "id": 468908, "author_name": "wakame", "author_url": "", "post_date": "2019-02-10T02:26:00.977000", "content": "

My best single model is CV 3.663 and LB 3.692 . Does anyone give me advice?

", "votes": 0, "replies": [] }, { "id": 468858, "author_name": "Manoj", "author_url": "", "post_date": "2019-02-09T21:35:10.957000", "content": "

So far i have come across best single model scoring LB 3.674 reported by @LongYin \nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73517#462946

", "votes": 0, "replies": [] }, { "id": 468780, "author_name": "FP", "author_url": "", "post_date": "2019-02-09T16:40:03.697000", "content": "

I can't even get past 3.687. So frustrating.

", "votes": 0, "replies": [ { "id": 468781, "author_name": "SubikashPal", "author_url": "", "post_date": "2019-02-09T16:43:56.103000", "content": "

use some sorts of aggregation based on Category. Should give some solid features.

", "votes": 2, "replies": [] } ] }, { "id": 468900, "author_name": "", "author_url": "", "post_date": "2019-02-10T01:37:48.003000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "468793": "my best model is at 3.688. just increase the diversity of your models :)", "468776": "I am struggling to improve my single model beyond 3.677. Any body who crossed the limit?", "472966": "wondering now", "472832": "I see you @SubikashPal. So you're starting to do some serious outlier prediction now huh? ", "472756": "I just got the 3.679 and wondering how could someone reaches to 3.670???", "468908": "My best single model is CV 3.663 and LB 3.692 . Does anyone give me advice?", "468858": "So far i have come across best single model scoring LB 3.674 reported by @LongYin \nhttps://www.kaggle.com/c/elo-merchant-category-recommendation/discussion/73517#462946", "468780": "I can't even get past 3.687. So frustrating.", "468900": "" } }