I was wondering if someone has tried this. Every site is unique in terms of weather pattern which drives energy consumption (besides building attributes). I am trying but the score is in range of 1.12-1.16. Any thoughts?
", "messages": [ { "id": "675451", "postDate": "11/18/2019 05:12:56", "content": "I was wondering if someone has tried this. Every site is unique in terms of weather pattern which drives energy consumption (besides building attributes). I am trying but the score is in range of 1.12-1.16. Any thoughts?
", "rawMarkdown": "I was wondering if someone has tried this. Every site is unique in terms of weather pattern which drives energy consumption (besides building attributes). I am trying but the score is in range of 1.12-1.16. Any thoughts?", "votes": null }, { "id": "675465", "postDate": "11/18/2019 05:28:38", "content": "I have shared a skeletal kernel that scores 1.15 (and with some minor changes, I've achieved 1.10 using the same snippet) without the use of any external data or labels.
", "rawMarkdown": "I have shared a skeletal [kernel](https://www.kaggle.com/rohanrao/ashrae-divide-and-conquer) that scores 1.15 (and with some minor changes, I've achieved 1.10 using the same snippet) without the use of any external data or labels.", "votes": null }, { "id": "675468", "postDate": "11/18/2019 05:36:00", "content": "I don't think it will make any difference for gbm models
", "rawMarkdown": "I don't think it will make any difference for gbm models", "votes": null }, { "id": "675480", "postDate": "11/18/2019 05:49:08", "content": "For the same set of parameters it won't make much of a difference, especially considering building_id is generally the top feature so LGBM starts with splitting on buildings (same for site-models). In fact, you'd expect the model on entire dataset to perform better.
\n\nBut, if you tune the hyper-parameters for each site-id model to optimally fit the data for that site-id, you can improve your score beyond 1.10
", "rawMarkdown": "For the same set of parameters it won't make much of a difference, especially considering building_id is generally the top feature so LGBM starts with splitting on buildings (same for site-models). In fact, you'd expect the model on entire dataset to perform better.\n\nBut, if you tune the hyper-parameters for each site-id model to optimally fit the data for that site-id, you can improve your score beyond 1.10", "votes": null }, { "id": "675493", "postDate": "11/18/2019 06:00:24", "content": "Thanks for the advice. Is your std between model scores is too high?
", "rawMarkdown": "Thanks for the advice. Is your std between model scores is too high?", "votes": null }, { "id": "675505", "postDate": "11/18/2019 06:26:48", "content": "@rohanrao @gunesevitan Thanks for the reply- this is helpful!
", "rawMarkdown": "rohanrao @gunesevitan Thanks for the reply- this is helpful!", "votes": null } ], "comments": [ { "id": 675465, "author_name": "rohanrao", "author_url": "", "post_date": "11/18/2019 05:28:38", "content": "I have shared a skeletal kernel that scores 1.15 (and with some minor changes, I've achieved 1.10 using the same snippet) without the use of any external data or labels.
", "votes": null, "replies": [ { "id": 675505, "author_name": "prasunmishra", "author_url": "", "post_date": "11/18/2019 06:26:48", "content": "@rohanrao @gunesevitan Thanks for the reply- this is helpful!
", "votes": null, "replies": [] } ] }, { "id": 675468, "author_name": "gunesevitan", "author_url": "", "post_date": "11/18/2019 05:36:00", "content": "I don't think it will make any difference for gbm models
", "votes": null, "replies": [ { "id": 675480, "author_name": "rohanrao", "author_url": "", "post_date": "11/18/2019 05:49:08", "content": "For the same set of parameters it won't make much of a difference, especially considering building_id is generally the top feature so LGBM starts with splitting on buildings (same for site-models). In fact, you'd expect the model on entire dataset to perform better.
\n\nBut, if you tune the hyper-parameters for each site-id model to optimally fit the data for that site-id, you can improve your score beyond 1.10
", "votes": null, "replies": [] }, { "id": 675493, "author_name": "gunesevitan", "author_url": "", "post_date": "11/18/2019 06:00:24", "content": "Thanks for the advice. Is your std between model scores is too high?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "675451": "I was wondering if someone has tried this. Every site is unique in terms of weather pattern which drives energy consumption (besides building attributes). I am trying but the score is in range of 1.12-1.16. Any thoughts?", "675465": "I have shared a skeletal [kernel](https://www.kaggle.com/rohanrao/ashrae-divide-and-conquer) that scores 1.15 (and with some minor changes, I've achieved 1.10 using the same snippet) without the use of any external data or labels.", "675468": "I don't think it will make any difference for gbm models", "675480": "For the same set of parameters it won't make much of a difference, especially considering building_id is generally the top feature so LGBM starts with splitting on buildings (same for site-models). In fact, you'd expect the model on entire dataset to perform better.\n\nBut, if you tune the hyper-parameters for each site-id model to optimally fit the data for that site-id, you can improve your score beyond 1.10", "675493": "Thanks for the advice. Is your std between model scores is too high?", "675505": "rohanrao @gunesevitan Thanks for the reply- this is helpful!" }, "source": "meta" }