There has been some discussions on how different sites and buildings behave differently and how timings might be misaligned across sites.\nHere is a Python kernel that builds models for each site individually. It need not specifically perform better but its CV and LB isn't all that bad (scores 1.15 on LB) and can help in ensembles. And who knows, it could even generalize well on private LB? 🤐
\n\nI can confirm that you can score 1.10 using site level models with some feature engineering and hyper-parameter tuning, so feel free to try out!
\n\nWould be interesting to know how others have fared with their site models (or maybe even building models!).
", "messages": [ { "id": 668433, "postDate": "2019-11-08T12:35:28.450Z", "content": "There has been some discussions on how different sites and buildings behave differently and how timings might be misaligned across sites.\nHere is a Python kernel that builds models for each site individually. It need not specifically perform better but its CV and LB isn't all that bad (scores 1.15 on LB) and can help in ensembles. And who knows, it could even generalize well on private LB? 🤐
\n\nI can confirm that you can score 1.10 using site level models with some feature engineering and hyper-parameter tuning, so feel free to try out!
\n\nWould be interesting to know how others have fared with their site models (or maybe even building models!).
", "rawMarkdown": "There has been some discussions on how different sites and buildings behave differently and how timings might be misaligned across sites.\nHere is a [Python kernel](https://www.kaggle.com/rohanrao/ashrae-divide-and-conquer) that builds models for each site individually. It need not specifically perform better but its CV and LB isn't all that bad (scores 1.15 on LB) and can help in ensembles. And who knows, it could even generalize well on private LB? 🤐\n\n* It might be able to capture site specific behaviour better\n* (To some extent) It gets enough information through lag features and hourly aggregates of misalignment features, assuming it is misaligned with train data as well as test data\n* It will lose out on information across sites\n* Current notebook uses the same parameters for each site which is not optimal\n* You could use different holiday dates for different sites\n\nI can confirm that you can score 1.10 using site level models with some feature engineering and hyper-parameter tuning, so feel free to try out!\n\nWould be interesting to know how others have fared with their site models (or maybe even building models!).", "votes": 29 }, { "id": 668671, "postDate": "2019-11-08T17:45:09.117Z", "content": "I built one. Takes forever and no improvement.
", "rawMarkdown": "I built one. Takes forever and no improvement.", "votes": 1 }, { "id": 668506, "postDate": "2019-11-08T13:58:56.667Z", "content": "Thank you for the information! What would you think of building a model for every building? Would it make sense to do that?
", "rawMarkdown": "Thank you for the information! What would you think of building a model for every building? Would it make sense to do that?", "votes": 1, "replies": [ { "id": 668964, "postDate": "2019-11-09T07:22:40.253Z", "content": "Building level is too granular and there isn't sufficient data points for a good stable model. My experiments with building level models didn't give good results so I dropped the idea.
", "rawMarkdown": "Building level is too granular and there isn't sufficient data points for a good stable model. My experiments with building level models didn't give good results so I dropped the idea.", "votes": 1 }, { "id": 669042, "postDate": "2019-11-09T11:37:22.650Z", "content": "Our team has built very simple building level models and it is performing decently on lb 1.12 but the only bottle neck is that we are out of features which can help us improve
", "rawMarkdown": "Our team has built very simple building level models and it is performing decently on lb 1.12 but the only bottle neck is that we are out of features which can help us improve", "votes": 2 } ] }, { "id": 681889, "postDate": "2019-11-26T16:39:58.617Z", "content": "In your code you are uploading leak files. Can you share those files? What was the source?
", "rawMarkdown": "In your code you are uploading leak files. Can you share those files? What was the source?", "replies": [ { "id": 681904, "postDate": "2019-11-26T16:59:04.540Z", "content": "Done. Those are just compiled from all the public kernels that share leak: Sites 0, 1, 2, 4, 15
", "rawMarkdown": "Done. Those are just compiled from all the public kernels that share leak: Sites 0, 1, 2, 4, 15" } ] }, { "id": 674521, "postDate": "2019-11-16T16:21:07.950Z", "content": "Got 1.12 LB score with site wise model, thanks a lot!
", "rawMarkdown": "Got 1.12 LB score with site wise model, thanks a lot! " }, { "id": 671544, "postDate": "2019-11-12T21:20:47.417Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 668671, "author_name": "Larry Schuster", "author_url": "", "post_date": "2019-11-08T17:45:09.117000", "content": "I built one. Takes forever and no improvement.
", "votes": 1, "replies": [] }, { "id": 668506, "author_name": "Carlo", "author_url": "", "post_date": "2019-11-08T13:58:56.667000", "content": "Thank you for the information! What would you think of building a model for every building? Would it make sense to do that?
", "votes": 1, "replies": [ { "id": 668964, "author_name": "Vopani", "author_url": "", "post_date": "2019-11-09T07:22:40.253000", "content": "Building level is too granular and there isn't sufficient data points for a good stable model. My experiments with building level models didn't give good results so I dropped the idea.
", "votes": 1, "replies": [] }, { "id": 669042, "author_name": "AdityaVikramSingh", "author_url": "", "post_date": "2019-11-09T11:37:22.650000", "content": "Our team has built very simple building level models and it is performing decently on lb 1.12 but the only bottle neck is that we are out of features which can help us improve
", "votes": 2, "replies": [] } ] }, { "id": 681889, "author_name": "Alexander Konstantinidis", "author_url": "", "post_date": "2019-11-26T16:39:58.617000", "content": "In your code you are uploading leak files. Can you share those files? What was the source?
", "votes": 0, "replies": [ { "id": 681904, "author_name": "Vopani", "author_url": "", "post_date": "2019-11-26T16:59:04.540000", "content": "Done. Those are just compiled from all the public kernels that share leak: Sites 0, 1, 2, 4, 15
", "votes": 0, "replies": [] } ] }, { "id": 674521, "author_name": "Yunxiao Song", "author_url": "", "post_date": "2019-11-16T16:21:07.950000", "content": "Got 1.12 LB score with site wise model, thanks a lot!
", "votes": 0, "replies": [] }, { "id": 671544, "author_name": "", "author_url": "", "post_date": "2019-11-12T21:20:47.417000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "668433": "There has been some discussions on how different sites and buildings behave differently and how timings might be misaligned across sites.\nHere is a [Python kernel](https://www.kaggle.com/rohanrao/ashrae-divide-and-conquer) that builds models for each site individually. It need not specifically perform better but its CV and LB isn't all that bad (scores 1.15 on LB) and can help in ensembles. And who knows, it could even generalize well on private LB? 🤐\n\n* It might be able to capture site specific behaviour better\n* (To some extent) It gets enough information through lag features and hourly aggregates of misalignment features, assuming it is misaligned with train data as well as test data\n* It will lose out on information across sites\n* Current notebook uses the same parameters for each site which is not optimal\n* You could use different holiday dates for different sites\n\nI can confirm that you can score 1.10 using site level models with some feature engineering and hyper-parameter tuning, so feel free to try out!\n\nWould be interesting to know how others have fared with their site models (or maybe even building models!).", "668671": "I built one. Takes forever and no improvement.", "668506": "Thank you for the information! What would you think of building a model for every building? Would it make sense to do that?", "681889": "In your code you are uploading leak files. Can you share those files? What was the source?", "674521": "Got 1.12 LB score with site wise model, thanks a lot! ", "671544": "" } }