{ "id": 116309, "title": "Best single model scores", "url": "/competitions/ashrae-energy-prediction/discussion/116309", "author_name": "Kostiantyn Isaienkov", "post_date": "2019-11-08T09:58:08.050000", "votes": 45, "comment_count": 51, "views": 0, "content": "

So blending has started and it is hard to understand from LB what who is using single model, who not. What is your best SINGLE model score?

\n\n

My results for now - 1.10 LB, 1.08 CV.

", "messages": [ { "id": 668342, "postDate": "2019-11-08T09:58:08.050Z", "content": "

So blending has started and it is hard to understand from LB what who is using single model, who not. What is your best SINGLE model score?

\n\n

My results for now - 1.10 LB, 1.08 CV.

", "rawMarkdown": "So blending has started and it is hard to understand from LB what who is using single model, who not. What is your best SINGLE model score?\n\nMy results for now - 1.10 LB, 1.08 CV.", "votes": 45 }, { "id": 676241, "postDate": "2019-11-19T03:29:43.853Z", "content": "

1.05 LB without any external data.\nI hope no more leakagešŸ˜„

", "rawMarkdown": "1.05 LB without any external data.\nI hope no more leakagešŸ˜„", "votes": 7 }, { "id": 668389, "postDate": "2019-11-08T11:26:02.410Z", "content": "

1.016 cv 1.07 lb

", "rawMarkdown": "1.016 cv 1.07 lb", "votes": 6, "replies": [ { "id": 668421, "postDate": "2019-11-08T12:15:11.200Z", "content": "

@oleg90777 , impressive result!\nAre you focused on FE or outlier detection strategy?

", "rawMarkdown": "@oleg90777 , impressive result!\nAre you focused on FE or outlier detection strategy?", "votes": 2 }, { "id": 668863, "postDate": "2019-11-09T01:08:40.613Z", "content": "

Your CV seems very close to LB. Are you using non-shuffle kfold ?

", "rawMarkdown": "Your CV seems very close to LB. Are you using non-shuffle kfold ?", "votes": 2 }, { "id": 668979, "postDate": "2019-11-09T08:10:45.580Z", "content": "

Yes

", "rawMarkdown": "Yes", "votes": 7 } ] }, { "id": 669586, "postDate": "2019-11-10T07:51:53.150Z", "content": "

0.998447158 cv 1.07 lb怂 but i don't know how to improve its up to 1.06;

", "rawMarkdown": "0.998447158 cv 1.07 lb怂 but i don't know how to improve its up to 1.06;", "votes": 4, "replies": [ { "id": 669910, "postDate": "2019-11-10T16:51:34.997Z", "content": "

@pp2file very impressive. Are you doing FE for weather data?

", "rawMarkdown": "@pp2file very impressive. Are you doing FE for weather data?" }, { "id": 670065, "postDate": "2019-11-11T00:21:21.693Z", "content": "

I test a lot weather FEļ¼Œbut it's not improve score. I just use normal weather. don't add lag.

", "rawMarkdown": "I test a lot weather FEļ¼Œbut it's not improve score. I just use normal weather. don't add lag.", "votes": 6 }, { "id": 671226, "postDate": "2019-11-12T12:00:00.827Z", "content": "

nice work!\nwhat is ur CV method?

", "rawMarkdown": "nice work!\nwhat is ur CV method?" }, { "id": 671278, "postDate": "2019-11-12T13:22:48.767Z", "content": "

just split two data. I think it's not a good method, Now i work to find a good cv method.

", "rawMarkdown": "just split two data. I think it's not a good method, Now i work to find a good cv method.", "votes": 6 }, { "id": 676383, "postDate": "2019-11-19T06:05:07.403Z", "content": "

@pp2file Congratulations on this jump :) , is there any cv method that you recommend from your side ?

", "rawMarkdown": "@pp2file Congratulations on this jump :) , is there any cv method that you recommend from your side ?" } ] }, { "id": 671044, "postDate": "2019-11-12T07:24:56.573Z", "content": "

1.07 CV, 1.09 LB

", "rawMarkdown": "1.07 CV, 1.09 LB", "votes": 1 }, { "id": 668942, "postDate": "2019-11-09T06:24:01.483Z", "content": "

I I suspect that the nature of this data means that single models won't end up being super effective. We'll probably have to combine a bunch of different models predicting different subsets to get best results.

\n\n

my best lb score is 1.10, but it's just the half-half kernel with the usual site_id 0 rows dropped, and a tiny bit of FE.

", "rawMarkdown": "I I suspect that the nature of this data means that single models won't end up being super effective. We'll probably have to combine a bunch of different models predicting different subsets to get best results. \n\nmy best lb score is 1.10, but it's just the half-half kernel with the usual site_id 0 rows dropped, and a tiny bit of FE. ", "votes": 1 }, { "id": 668403, "postDate": "2019-11-08T11:43:19.680Z", "content": "

1.10LB

", "rawMarkdown": "1.10LB", "votes": 1 }, { "id": 668395, "postDate": "2019-11-08T11:30:59.927Z", "content": "

1.05 cv 1.09 lb

", "rawMarkdown": "1.05 cv 1.09 lb", "votes": 1 }, { "id": 683468, "postDate": "2019-11-28T13:11:43.333Z", "content": "

1.07 LB without any leak

", "rawMarkdown": "1.07 LB without any leak", "votes": 2 }, { "id": 682501, "postDate": "2019-11-27T14:16:29.340Z", "content": "

Without any leak, using Catboost 5 fold, lb = 1.06 and cv = 1.2073

", "rawMarkdown": "Without any leak, using Catboost 5 fold, lb = 1.06 and cv = 1.2073", "votes": 2 }, { "id": 668485, "postDate": "2019-11-08T13:25:40.263Z", "content": "

1.07 lb

", "rawMarkdown": "1.07 lb ", "votes": 2, "replies": [ { "id": 668486, "postDate": "2019-11-08T13:30:36.533Z", "content": "

Hi, @yamsam , perfect!\nThe same question, are you focused on FE or outlier detection strategy?

", "rawMarkdown": "Hi, @yamsam , perfect!\nThe same question, are you focused on FE or outlier detection strategy?" }, { "id": 668501, "postDate": "2019-11-08T13:49:40.897Z", "content": "

I have no magic feature and don't use ucf data yet

", "rawMarkdown": "I have no magic feature and don't use ucf data yet", "votes": 1 }, { "id": 668943, "postDate": "2019-11-09T06:29:12.847Z", "content": "

dang, you must be pretty good at this

", "rawMarkdown": "dang, you must be pretty good at this" }, { "id": 669018, "postDate": "2019-11-09T10:06:24.677Z", "content": "

sorry for stupid question, what is ucf data ?

", "rawMarkdown": "sorry for stupid question, what is ucf data ?", "votes": 2 }, { "id": 669072, "postDate": "2019-11-09T12:47:40.193Z", "content": "

https://www.oeis.ucf.edu/buildings

", "rawMarkdown": "https://www.oeis.ucf.edu/buildings", "votes": 1 }, { "id": 673635, "postDate": "2019-11-15T08:53:22.860Z", "content": "

Update. LB 1.03 with UCF data

", "rawMarkdown": "Update. LB 1.03 with UCF data" }, { "id": 673745, "postDate": "2019-11-15T12:32:02.343Z", "content": "

my LB 1.06 I am modifing your notebook with UCF

", "rawMarkdown": "my LB 1.06 I am modifing your notebook with UCF", "votes": 1 }, { "id": 673837, "postDate": "2019-11-15T14:42:20.727Z", "content": "

Isn't the ucf data kind of already providing the target variable results ?

", "rawMarkdown": "Isn't the ucf data kind of already providing the target variable results ?" }, { "id": 674180, "postDate": "2019-11-16T01:02:56.113Z", "content": "

Yes. but take care, next wave is incoming.

", "rawMarkdown": "Yes. but take care, next wave is incoming." }, { "id": 674273, "postDate": "2019-11-16T06:19:15.280Z", "content": "

@yamsam Using UCF data(site 0 ) my cv and lb score is increasing by approx 0.01. Also i am splitting train /validation(UCF data). Are you training with given data and making change in submission file by replacing directly site 0 meter reading ?

", "rawMarkdown": "@yamsam Using UCF data(site 0 ) my cv and lb score is increasing by approx 0.01. Also i am splitting train /validation(UCF data). Are you training with given data and making change in submission file by replacing directly site 0 meter reading ? " }, { "id": 674276, "postDate": "2019-11-16T06:34:01.350Z", "content": "

YES. But I have not found the best CV with UCF data. I am just trying to find it in my public kernel.

", "rawMarkdown": "YES. But I have not found the best CV with UCF data. I am just trying to find it in my public kernel.", "votes": 1 }, { "id": 674289, "postDate": "2019-11-16T07:08:13.650Z", "content": "

Thank you for clarification.

", "rawMarkdown": "Thank you for clarification." } ] }, { "id": 675453, "postDate": "2019-11-18T05:13:51.137Z", "content": "

Built model by meter type using 5 KFold (shuffle = False),using:\nhttps://www.kaggle.com/nz0722/aligned-timestamp-lgbm-by-meter-type\ncv = 1.2406, lb = 1.04

\n\n

P.s : Including leaked data too

", "rawMarkdown": "Built model by meter type using 5 KFold (shuffle = False),using:\nhttps://www.kaggle.com/nz0722/aligned-timestamp-lgbm-by-meter-type\ncv = 1.2406, lb = 1.04\n\nP.s : Including leaked data too\n", "votes": 1, "replies": [ { "id": 675481, "postDate": "2019-11-18T05:49:09.943Z", "content": "

NO UCF DATA ?

", "rawMarkdown": "NO UCF DATA ?" }, { "id": 675486, "postDate": "2019-11-18T05:49:58.353Z", "content": "

without leaked data? It is impressive.

", "rawMarkdown": "without leaked data? It is impressive." }, { "id": 675503, "postDate": "2019-11-18T06:24:41.277Z", "content": "

Sorry for the confusion. I used leaked data

", "rawMarkdown": "Sorry for the confusion. I used leaked data", "votes": 1 } ] }, { "id": 684787, "postDate": "2019-11-30T13:20:10.253Z", "content": "

Very helpful thread

", "rawMarkdown": "Very helpful thread" }, { "id": 683952, "postDate": "2019-11-29T04:10:07.427Z", "content": "

LB1.06 CV0.81 I donā€™t known why the cv score is so low

", "rawMarkdown": "LB1.06 CV0.81 I donā€™t known why the cv score is so low\n", "replies": [ { "id": 684055, "postDate": "2019-11-29T07:12:01.523Z", "content": "

your model is overfit

", "rawMarkdown": "your model is overfit" } ] }, { "id": 679497, "postDate": "2019-11-22T21:37:12.500Z", "content": "

LB 1.10 , hyperparametes tuning helped to gain 0.01, I'll keep searching for the optimal ones

", "rawMarkdown": "LB 1.10 , hyperparametes tuning helped to gain 0.01, I'll keep searching for the optimal ones" }, { "id": 675337, "postDate": "2019-11-18T00:23:57.583Z", "content": "

1.09LB

", "rawMarkdown": "1.09LB" }, { "id": 671427, "postDate": "2019-11-12T17:21:43.943Z", "content": "

1.0015554 cv 1.07 lb no ucf data

", "rawMarkdown": "1.0015554 cv 1.07 lb no ucf data", "replies": [ { "id": 671442, "postDate": "2019-11-12T17:52:25.443Z", "content": "

@negi009 what's the secret sauceā˜€ļø šŸ˜œ , some sophisticated FE or using normal data?

", "rawMarkdown": "@negi009 what's the secret sauceā˜€ļø šŸ˜œ , some sophisticated FE or using normal data?" }, { "id": 671573, "postDate": "2019-11-12T22:17:09.327Z", "content": "

What CV are you referring to?

", "rawMarkdown": "What CV are you referring to?" }, { "id": 672065, "postDate": "2019-11-13T13:34:01.783Z", "content": "

normal data very few FE # how you split data can give you better model.

", "rawMarkdown": "normal data very few FE # how you split data can give you better model. ", "votes": 2 }, { "id": 676384, "postDate": "2019-11-19T06:06:01.050Z", "content": "

@negi009 do you recommend a time based split for various meters ?

", "rawMarkdown": "@negi009 do you recommend a time based split for various meters ?\n" }, { "id": 676426, "postDate": "2019-11-19T07:03:47.803Z", "content": "

@avikrams yeah split should be time based to avoid leakage.

", "rawMarkdown": "@avikrams yeah split should be time based to avoid leakage. ", "votes": 1 } ] }, { "id": 669907, "postDate": "2019-11-10T16:48:06.813Z", "content": "

1.09 LB

", "rawMarkdown": "1.09 LB" }, { "id": 668524, "postDate": "2019-11-08T14:21:29.817Z", "content": "

1.09 LB

", "rawMarkdown": "1.09 LB" }, { "id": 668502, "postDate": "2019-11-08T13:50:12.543Z", "content": "

1.08LB

", "rawMarkdown": "1.08LB", "replies": [ { "id": 669024, "postDate": "2019-11-09T10:44:59.410Z", "content": "

that is amazing. i have tried but its hard to get better from 1.1LB

", "rawMarkdown": "that is amazing. i have tried but its hard to get better from 1.1LB", "votes": 1 } ] }, { "id": 672456, "postDate": "2019-11-13T23:10:34.693Z", "rawMarkdown": "", "isDeleted": true }, { "id": 668754, "postDate": "2019-11-08T19:58:53.840Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 676241, "author_name": "Hiroyuki Namba", "author_url": "", "post_date": "2019-11-19T03:29:43.853000", "content": "

1.05 LB without any external data.\nI hope no more leakagešŸ˜„

", "votes": 7, "replies": [] }, { "id": 668389, "author_name": "Oleg Knaub", "author_url": "", "post_date": "2019-11-08T11:26:02.410000", "content": "

1.016 cv 1.07 lb

", "votes": 6, "replies": [ { "id": 668421, "author_name": "Kostiantyn Isaienkov", "author_url": "", "post_date": "2019-11-08T12:15:11.200000", "content": "

@oleg90777 , impressive result!\nAre you focused on FE or outlier detection strategy?

", "votes": 2, "replies": [] }, { "id": 668863, "author_name": "Antoine", "author_url": "", "post_date": "2019-11-09T01:08:40.613000", "content": "

Your CV seems very close to LB. Are you using non-shuffle kfold ?

", "votes": 2, "replies": [] }, { "id": 668979, "author_name": "Oleg Knaub", "author_url": "", "post_date": "2019-11-09T08:10:45.580000", "content": "

Yes

", "votes": 7, "replies": [] } ] }, { "id": 669586, "author_name": "Zhang Yunfei", "author_url": "", "post_date": "2019-11-10T07:51:53.150000", "content": "

0.998447158 cv 1.07 lb怂 but i don't know how to improve its up to 1.06;

", "votes": 4, "replies": [ { "id": 669910, "author_name": "Vishal Sharma", "author_url": "", "post_date": "2019-11-10T16:51:34.997000", "content": "

@pp2file very impressive. Are you doing FE for weather data?

", "votes": 0, "replies": [] }, { "id": 670065, "author_name": "Zhang Yunfei", "author_url": "", "post_date": "2019-11-11T00:21:21.693000", "content": "

I test a lot weather FEļ¼Œbut it's not improve score. I just use normal weather. don't add lag.

", "votes": 6, "replies": [] }, { "id": 671226, "author_name": "shadow-boxing", "author_url": "", "post_date": "2019-11-12T12:00:00.827000", "content": "

nice work!\nwhat is ur CV method?

", "votes": 0, "replies": [] }, { "id": 671278, "author_name": "Zhang Yunfei", "author_url": "", "post_date": "2019-11-12T13:22:48.767000", "content": "

just split two data. I think it's not a good method, Now i work to find a good cv method.

", "votes": 6, "replies": [] }, { "id": 676383, "author_name": "AdityaVikramSingh", "author_url": "", "post_date": "2019-11-19T06:05:07.403000", "content": "

@pp2file Congratulations on this jump :) , is there any cv method that you recommend from your side ?

", "votes": 0, "replies": [] } ] }, { "id": 671044, "author_name": "Chau Huynh", "author_url": "", "post_date": "2019-11-12T07:24:56.573000", "content": "

1.07 CV, 1.09 LB

", "votes": 1, "replies": [] }, { "id": 668942, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2019-11-09T06:24:01.483000", "content": "

I I suspect that the nature of this data means that single models won't end up being super effective. We'll probably have to combine a bunch of different models predicting different subsets to get best results.

\n\n

my best lb score is 1.10, but it's just the half-half kernel with the usual site_id 0 rows dropped, and a tiny bit of FE.

", "votes": 1, "replies": [] }, { "id": 668403, "author_name": "Jie Lu", "author_url": "", "post_date": "2019-11-08T11:43:19.680000", "content": "

1.10LB

", "votes": 1, "replies": [] }, { "id": 668395, "author_name": "Davide Stenner", "author_url": "", "post_date": "2019-11-08T11:30:59.927000", "content": "

1.05 cv 1.09 lb

", "votes": 1, "replies": [] }, { "id": 683468, "author_name": "", "author_url": "", "post_date": "2019-11-28T13:11:43.333000", "content": "

1.07 LB without any leak

", "votes": 2, "replies": [] }, { "id": 682501, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-11-27T14:16:29.340000", "content": "

Without any leak, using Catboost 5 fold, lb = 1.06 and cv = 1.2073

", "votes": 2, "replies": [] }, { "id": 668485, "author_name": "Isamu", "author_url": "", "post_date": "2019-11-08T13:25:40.263000", "content": "

1.07 lb

", "votes": 2, "replies": [ { "id": 668486, "author_name": "Kostiantyn Isaienkov", "author_url": "", "post_date": "2019-11-08T13:30:36.533000", "content": "

Hi, @yamsam , perfect!\nThe same question, are you focused on FE or outlier detection strategy?

", "votes": 0, "replies": [] }, { "id": 668501, "author_name": "Isamu", "author_url": "", "post_date": "2019-11-08T13:49:40.897000", "content": "

I have no magic feature and don't use ucf data yet

", "votes": 1, "replies": [] }, { "id": 668943, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2019-11-09T06:29:12.847000", "content": "

dang, you must be pretty good at this

", "votes": 0, "replies": [] }, { "id": 669018, "author_name": "Jonas Matuzas", "author_url": "", "post_date": "2019-11-09T10:06:24.677000", "content": "

sorry for stupid question, what is ucf data ?

", "votes": 2, "replies": [] }, { "id": 669072, "author_name": "Isamu", "author_url": "", "post_date": "2019-11-09T12:47:40.193000", "content": "

https://www.oeis.ucf.edu/buildings

", "votes": 1, "replies": [] }, { "id": 673635, "author_name": "Isamu", "author_url": "", "post_date": "2019-11-15T08:53:22.860000", "content": "

Update. LB 1.03 with UCF data

", "votes": 0, "replies": [] }, { "id": 673745, "author_name": "Jonas Matuzas", "author_url": "", "post_date": "2019-11-15T12:32:02.343000", "content": "

my LB 1.06 I am modifing your notebook with UCF

", "votes": 1, "replies": [] }, { "id": 673837, "author_name": "DietHard", "author_url": "", "post_date": "2019-11-15T14:42:20.727000", "content": "

Isn't the ucf data kind of already providing the target variable results ?

", "votes": 0, "replies": [] }, { "id": 674180, "author_name": "Isamu", "author_url": "", "post_date": "2019-11-16T01:02:56.113000", "content": "

Yes. but take care, next wave is incoming.

", "votes": 0, "replies": [] }, { "id": 674273, "author_name": "ADITYA KUMAR (Negi)", "author_url": "", "post_date": "2019-11-16T06:19:15.280000", "content": "

@yamsam Using UCF data(site 0 ) my cv and lb score is increasing by approx 0.01. Also i am splitting train /validation(UCF data). Are you training with given data and making change in submission file by replacing directly site 0 meter reading ?

", "votes": 0, "replies": [] }, { "id": 674276, "author_name": "Isamu", "author_url": "", "post_date": "2019-11-16T06:34:01.350000", "content": "

YES. But I have not found the best CV with UCF data. I am just trying to find it in my public kernel.

", "votes": 1, "replies": [] }, { "id": 674289, "author_name": "ADITYA KUMAR (Negi)", "author_url": "", "post_date": "2019-11-16T07:08:13.650000", "content": "

Thank you for clarification.

", "votes": 0, "replies": [] } ] }, { "id": 675453, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-11-18T05:13:51.137000", "content": "

Built model by meter type using 5 KFold (shuffle = False),using:\nhttps://www.kaggle.com/nz0722/aligned-timestamp-lgbm-by-meter-type\ncv = 1.2406, lb = 1.04

\n\n

P.s : Including leaked data too

", "votes": 1, "replies": [ { "id": 675481, "author_name": "ADITYA KUMAR (Negi)", "author_url": "", "post_date": "2019-11-18T05:49:09.943000", "content": "

NO UCF DATA ?

", "votes": 0, "replies": [] }, { "id": 675486, "author_name": "Johnny Lee", "author_url": "", "post_date": "2019-11-18T05:49:58.353000", "content": "

without leaked data? It is impressive.

", "votes": 0, "replies": [] }, { "id": 675503, "author_name": "HarshitMehta", "author_url": "", "post_date": "2019-11-18T06:24:41.277000", "content": "

Sorry for the confusion. I used leaked data

", "votes": 1, "replies": [] } ] }, { "id": 684787, "author_name": "NikhilGupta", "author_url": "", "post_date": "2019-11-30T13:20:10.253000", "content": "

Very helpful thread

", "votes": 0, "replies": [] }, { "id": 683952, "author_name": "Dxx", "author_url": "", "post_date": "2019-11-29T04:10:07.427000", "content": "

LB1.06 CV0.81 I donā€™t known why the cv score is so low

", "votes": 0, "replies": [ { "id": 684055, "author_name": "", "author_url": "", "post_date": "2019-11-29T07:12:01.523000", "content": "

your model is overfit

", "votes": 0, "replies": [] } ] }, { "id": 679497, "author_name": "Igor Dudchenko", "author_url": "", "post_date": "2019-11-22T21:37:12.500000", "content": "

LB 1.10 , hyperparametes tuning helped to gain 0.01, I'll keep searching for the optimal ones

", "votes": 0, "replies": [] }, { "id": 675337, "author_name": "Shahules", "author_url": "", "post_date": "2019-11-18T00:23:57.583000", "content": "

1.09LB

", "votes": 0, "replies": [] }, { "id": 671427, "author_name": "ADITYA KUMAR (Negi)", "author_url": "", "post_date": "2019-11-12T17:21:43.943000", "content": "

1.0015554 cv 1.07 lb no ucf data

", "votes": 0, "replies": [ { "id": 671442, "author_name": "Aadil Srivastava", "author_url": "", "post_date": "2019-11-12T17:52:25.443000", "content": "

@negi009 what's the secret sauceā˜€ļø šŸ˜œ , some sophisticated FE or using normal data?

", "votes": 0, "replies": [] }, { "id": 671573, "author_name": "Antoine", "author_url": "", "post_date": "2019-11-12T22:17:09.327000", "content": "

What CV are you referring to?

", "votes": 0, "replies": [] }, { "id": 672065, "author_name": "ADITYA KUMAR (Negi)", "author_url": "", "post_date": "2019-11-13T13:34:01.783000", "content": "

normal data very few FE # how you split data can give you better model.

", "votes": 2, "replies": [] }, { "id": 676384, "author_name": "AdityaVikramSingh", "author_url": "", "post_date": "2019-11-19T06:06:01.050000", "content": "

@negi009 do you recommend a time based split for various meters ?

", "votes": 0, "replies": [] }, { "id": 676426, "author_name": "ADITYA KUMAR (Negi)", "author_url": "", "post_date": "2019-11-19T07:03:47.803000", "content": "

@avikrams yeah split should be time based to avoid leakage.

", "votes": 1, "replies": [] } ] }, { "id": 669907, "author_name": "Vishal Sharma", "author_url": "", "post_date": "2019-11-10T16:48:06.813000", "content": "

1.09 LB

", "votes": 0, "replies": [] }, { "id": 668524, "author_name": "Kangyu Chen", "author_url": "", "post_date": "2019-11-08T14:21:29.817000", "content": "

1.09 LB

", "votes": 0, "replies": [] }, { "id": 668502, "author_name": "Mr Loke", "author_url": "", "post_date": "2019-11-08T13:50:12.543000", "content": "

1.08LB

", "votes": 0, "replies": [ { "id": 669024, "author_name": "Yixinchen", "author_url": "", "post_date": "2019-11-09T10:44:59.410000", "content": "

that is amazing. i have tried but its hard to get better from 1.1LB

", "votes": 1, "replies": [] } ] }, { "id": 672456, "author_name": "", "author_url": "", "post_date": "2019-11-13T23:10:34.693000", "content": "", "votes": 0, "replies": [] }, { "id": 668754, "author_name": "", "author_url": "", "post_date": "2019-11-08T19:58:53.840000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "668342": "So blending has started and it is hard to understand from LB what who is using single model, who not. What is your best SINGLE model score?\n\nMy results for now - 1.10 LB, 1.08 CV.", "676241": "1.05 LB without any external data.\nI hope no more leakagešŸ˜„", "668389": "1.016 cv 1.07 lb", "669586": "0.998447158 cv 1.07 lb怂 but i don't know how to improve its up to 1.06;", "671044": "1.07 CV, 1.09 LB", "668942": "I I suspect that the nature of this data means that single models won't end up being super effective. We'll probably have to combine a bunch of different models predicting different subsets to get best results. \n\nmy best lb score is 1.10, but it's just the half-half kernel with the usual site_id 0 rows dropped, and a tiny bit of FE. ", "668403": "1.10LB", "668395": "1.05 cv 1.09 lb", "683468": "1.07 LB without any leak", "682501": "Without any leak, using Catboost 5 fold, lb = 1.06 and cv = 1.2073", "668485": "1.07 lb ", "675453": "Built model by meter type using 5 KFold (shuffle = False),using:\nhttps://www.kaggle.com/nz0722/aligned-timestamp-lgbm-by-meter-type\ncv = 1.2406, lb = 1.04\n\nP.s : Including leaked data too\n", "684787": "Very helpful thread", "683952": "LB1.06 CV0.81 I donā€™t known why the cv score is so low\n", "679497": "LB 1.10 , hyperparametes tuning helped to gain 0.01, I'll keep searching for the optimal ones", "675337": "1.09LB", "671427": "1.0015554 cv 1.07 lb no ucf data", "669907": "1.09 LB", "668524": "1.09 LB", "668502": "1.08LB", "672456": "", "668754": "" } }