Let's assume for the moment blending different lgboost models is not the only possible way forward. Has anyone tried treating the problem as a time series forecasting task? Lumping everything together obviously won't work - I am currently trying to decide what is the right aggregation level where it starts making sense: per building? building x meter?
\n\nCurious to hear what others think.
", "messages": [ { "id": 667204, "postDate": "2019-11-07T00:10:04.260Z", "content": "Let's assume for the moment blending different lgboost models is not the only possible way forward. Has anyone tried treating the problem as a time series forecasting task? Lumping everything together obviously won't work - I am currently trying to decide what is the right aggregation level where it starts making sense: per building? building x meter?
\n\nCurious to hear what others think.
", "rawMarkdown": "Let's assume for the moment blending different lgboost models is not the only possible way forward. Has anyone tried treating the problem as a time series forecasting task? Lumping everything together obviously won't work - I am currently trying to decide what is the right aggregation level where it starts making sense: per building? building x meter? \n\nCurious to hear what others think.\n", "votes": 12 }, { "id": 667392, "postDate": "2019-11-07T07:24:00.077Z", "content": "The energy aspects are so different among them, that an aggregation level per building has no sense.
\n\nWithout a doubt, building x meter.
", "rawMarkdown": "The energy aspects are so different among them, that an aggregation level per building has no sense.\n\nWithout a doubt, building x meter.", "votes": 1 }, { "id": 668415, "postDate": "2019-11-08T12:04:22.290Z", "content": "It is an interesting question. It seems natural to treat it as a time series problem. However, predicting two years ahead with only one year of data is very hard to bootstrap and new features can easily encode noise. Removing the time aspect may or may not be helpful.
\n\nCurious to see what the final winning solution will be. There seem to be a lot of hidden treasures inside the data.
", "rawMarkdown": "It is an interesting question. It seems natural to treat it as a time series problem. However, predicting two years ahead with only one year of data is very hard to bootstrap and new features can easily encode noise. Removing the time aspect may or may not be helpful.\n\nCurious to see what the final winning solution will be. There seem to be a lot of hidden treasures inside the data.", "votes": 2, "replies": [ { "id": 668811, "postDate": "2019-11-08T22:59:52.707Z", "content": "there s a lot of (not so well) hidden trash in the data
", "rawMarkdown": "there s a lot of (not so well) hidden trash in the data", "votes": 4 }, { "id": 668820, "postDate": "2019-11-08T23:24:48.483Z", "content": "Haha, agree!
", "rawMarkdown": "Haha, agree! " }, { "id": 674089, "postDate": "2019-11-15T21:58:27.820Z", "content": "Now that we have some of the locations known perhaps we could recover more than one years worth of data to train on.
", "rawMarkdown": "Now that we have some of the locations known perhaps we could recover more than one years worth of data to train on." } ] }, { "id": 667887, "postDate": "2019-11-07T19:48:05.813Z", "content": "I remember that in the WTF competition my Arima model with external regressors was pretty competitive. It would be interesting to create a similar model with seasonal components.
", "rawMarkdown": "I remember that in the WTF competition my Arima model with external regressors was pretty competitive. It would be interesting to create a similar model with seasonal components.", "votes": 2, "replies": [ { "id": 667915, "postDate": "2019-11-07T20:13:42.120Z", "content": "Seriously there has been a WTF competition?? How come I missed that??
", "rawMarkdown": "Seriously there has been a WTF competition?? How come I missed that??" }, { "id": 667919, "postDate": "2019-11-07T20:18:48.763Z", "content": "Heh, that was an acronym for Wikipedia Traffic Forecasting.
", "rawMarkdown": "Heh, that was an acronym for [Wikipedia Traffic Forecasting](https://www.kaggle.com/c/web-traffic-time-series-forecasting).", "votes": 1 }, { "id": 667997, "postDate": "2019-11-07T21:29:08.017Z", "content": "@kailex Would ARIMA work for two years though?
", "rawMarkdown": "@kailex Would ARIMA work for two years though? " }, { "id": 668018, "postDate": "2019-11-07T21:55:28.547Z", "content": "It might work with proper tuning, regressors and Fourier components.
", "rawMarkdown": "It might work with proper tuning, regressors and Fourier components." } ] }, { "id": 667216, "postDate": "2019-11-07T00:48:42.660Z", "content": "What about partial Fourier decomposition and regolarization with baseline aggregation by (site_id, weekday, and primary_use)?
\n\nUnfortunately I have too little time for these competitions..especially if most of the work has to deal with cleaning messy data.
", "rawMarkdown": "What about partial Fourier decomposition and regolarization with baseline aggregation by (site_id, weekday, and primary_use)? \n\nUnfortunately I have too little time for these competitions..especially if most of the work has to deal with cleaning messy data. \n ", "replies": [ { "id": 667389, "postDate": "2019-11-07T07:19:01.367Z", "content": "eww that sounds like math
", "rawMarkdown": "eww that sounds like math" } ] }, { "id": 667219, "postDate": "2019-11-07T01:00:42.523Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 667392, "author_name": "Juanma Hernández", "author_url": "", "post_date": "2019-11-07T07:24:00.077000", "content": "The energy aspects are so different among them, that an aggregation level per building has no sense.
\n\nWithout a doubt, building x meter.
", "votes": 1, "replies": [] }, { "id": 668415, "author_name": "Carlo", "author_url": "", "post_date": "2019-11-08T12:04:22.290000", "content": "It is an interesting question. It seems natural to treat it as a time series problem. However, predicting two years ahead with only one year of data is very hard to bootstrap and new features can easily encode noise. Removing the time aspect may or may not be helpful.
\n\nCurious to see what the final winning solution will be. There seem to be a lot of hidden treasures inside the data.
", "votes": 2, "replies": [ { "id": 668811, "author_name": "andrea cazzaniga", "author_url": "", "post_date": "2019-11-08T22:59:52.707000", "content": "there s a lot of (not so well) hidden trash in the data
", "votes": 4, "replies": [] }, { "id": 668820, "author_name": "Carlo", "author_url": "", "post_date": "2019-11-08T23:24:48.483000", "content": "Haha, agree!
", "votes": 0, "replies": [] }, { "id": 674089, "author_name": "Rob Rose", "author_url": "", "post_date": "2019-11-15T21:58:27.820000", "content": "Now that we have some of the locations known perhaps we could recover more than one years worth of data to train on.
", "votes": 0, "replies": [] } ] }, { "id": 667887, "author_name": "kxx", "author_url": "", "post_date": "2019-11-07T19:48:05.813000", "content": "I remember that in the WTF competition my Arima model with external regressors was pretty competitive. It would be interesting to create a similar model with seasonal components.
", "votes": 2, "replies": [ { "id": 667915, "author_name": "bluetrain", "author_url": "", "post_date": "2019-11-07T20:13:42.120000", "content": "Seriously there has been a WTF competition?? How come I missed that??
", "votes": 0, "replies": [] }, { "id": 667919, "author_name": "kxx", "author_url": "", "post_date": "2019-11-07T20:18:48.763000", "content": "Heh, that was an acronym for Wikipedia Traffic Forecasting.
", "votes": 1, "replies": [] }, { "id": 667997, "author_name": "Tonghui Li", "author_url": "", "post_date": "2019-11-07T21:29:08.017000", "content": "@kailex Would ARIMA work for two years though?
", "votes": 0, "replies": [] }, { "id": 668018, "author_name": "kxx", "author_url": "", "post_date": "2019-11-07T21:55:28.547000", "content": "It might work with proper tuning, regressors and Fourier components.
", "votes": 0, "replies": [] } ] }, { "id": 667216, "author_name": "andrea cazzaniga", "author_url": "", "post_date": "2019-11-07T00:48:42.660000", "content": "What about partial Fourier decomposition and regolarization with baseline aggregation by (site_id, weekday, and primary_use)?
\n\nUnfortunately I have too little time for these competitions..especially if most of the work has to deal with cleaning messy data.
", "votes": 0, "replies": [ { "id": 667389, "author_name": "Louka Ewington-Pitsos", "author_url": "", "post_date": "2019-11-07T07:19:01.367000", "content": "eww that sounds like math
", "votes": 0, "replies": [] } ] }, { "id": 667219, "author_name": "", "author_url": "", "post_date": "2019-11-07T01:00:42.523000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "667204": "Let's assume for the moment blending different lgboost models is not the only possible way forward. Has anyone tried treating the problem as a time series forecasting task? Lumping everything together obviously won't work - I am currently trying to decide what is the right aggregation level where it starts making sense: per building? building x meter? \n\nCurious to hear what others think.\n", "667392": "The energy aspects are so different among them, that an aggregation level per building has no sense.\n\nWithout a doubt, building x meter.", "668415": "It is an interesting question. It seems natural to treat it as a time series problem. However, predicting two years ahead with only one year of data is very hard to bootstrap and new features can easily encode noise. Removing the time aspect may or may not be helpful.\n\nCurious to see what the final winning solution will be. There seem to be a lot of hidden treasures inside the data.", "667887": "I remember that in the WTF competition my Arima model with external regressors was pretty competitive. It would be interesting to create a similar model with seasonal components.", "667216": "What about partial Fourier decomposition and regolarization with baseline aggregation by (site_id, weekday, and primary_use)? \n\nUnfortunately I have too little time for these competitions..especially if most of the work has to deal with cleaning messy data. \n ", "667219": "" } }