Do I understand correctly that we need to prediction the 2015-2019 seasons? In order to avoid leakage, we should the 2015-2019 seasons removed from the training dataset and then it is worth building a model that best predicts these seasons, according data from 1985-2014 seasons. Is that right?
\n\nSee my kernel: MM NCAAM [No leaks]: LGB, XGB, LogReg
\n\nAlso answers there are in the dataset NCAA 2015-2019 labels from @nroman (really played matches of 2015-2019).
\n\nOr in the post \"How to score 0 and why you should not do this\" from @nroman, there is an instruction on exactly how to read it data from data of the competition.
\n\nDid I understand everything correctly?
", "messages": [ { "id": "765476", "postDate": "03/06/2020 17:00:49", "content": "Do I understand correctly that we need to prediction the 2015-2019 seasons? In order to avoid leakage, we should the 2015-2019 seasons removed from the training dataset and then it is worth building a model that best predicts these seasons, according data from 1985-2014 seasons. Is that right?
\n\nSee my kernel: MM NCAAM [No leaks]: LGB, XGB, LogReg
\n\nAlso answers there are in the dataset NCAA 2015-2019 labels from @nroman (really played matches of 2015-2019).
\n\nOr in the post \"How to score 0 and why you should not do this\" from @nroman, there is an instruction on exactly how to read it data from data of the competition.
\n\nDid I understand everything correctly?
", "rawMarkdown": "Do I understand correctly that we need to prediction the 2015-2019 seasons? In order to avoid leakage, we should the 2015-2019 seasons removed from the training dataset and then it is worth building a model that best predicts these seasons, according data from 1985-2014 seasons. Is that right?\n\nSee my kernel: [MM NCAAM [No leaks]: LGB, XGB, LogReg](https://www.kaggle.com/vbmokin/mm-ncaam-no-leaks-lgb-xgb-logreg)\n\nAlso answers there are in the dataset [NCAA 2015-2019 labels](https://www.kaggle.com/nroman/ncaa-20152019-labels) from @nroman (really played matches of 2015-2019). \n\nOr in the post \"[How to score 0 and why you should not do this](https://www.kaggle.com/c/google-cloud-ncaa-march-madness-2020-division-1-mens-tournament/discussion/131539)\" from @nroman, there is an instruction on exactly how to read it data from data of the competition. \n\nDid I understand everything correctly?", "votes": null }, { "id": "765500", "postDate": "03/06/2020 17:52:20", "content": "Yes, stage 1 is in place primarily to give users time to understand the format of the data and test their workflows for building models.
\n\nWe will only have a few days to incorporate the new data provided in stage 2 and predict the answers for this years tournament. So you will almost certainly want to retrain your model in stage 2 using the new data, but stage 1 allows you to test your methods so that you can do it quickly in stage 2.
", "rawMarkdown": "Yes, stage 1 is in place primarily to give users time to understand the format of the data and test their workflows for building models.\n\nWe will only have a few days to incorporate the new data provided in stage 2 and predict the answers for this years tournament. So you will almost certainly want to retrain your model in stage 2 using the new data, but stage 1 allows you to test your methods so that you can do it quickly in stage 2.", "votes": null }, { "id": "765505", "postDate": "03/06/2020 17:57:43", "content": "Basically yes, you got the idea. Stage one is mostly to play around. I have posted true labels so people can validate their models offline, without need to submit every time.
", "rawMarkdown": "Basically yes, you got the idea. Stage one is mostly to play around. I have posted true labels so people can validate their models offline, without need to submit every time.", "votes": null }, { "id": "765719", "postDate": "03/07/2020 01:55:57", "content": "Yes, that's good.But without removing the entire 2015-2019 data from the train you can use data till 2016 to train and predict for 2017 and vice versa.This is a much better approach..
", "rawMarkdown": "Yes, that's good.But without removing the entire 2015-2019 data from the train you can use data till 2016 to train and predict for 2017 and vice versa.This is a much better approach..", "votes": null } ], "comments": [ { "id": 765500, "author_name": "rsa013", "author_url": "", "post_date": "03/06/2020 17:52:20", "content": "Yes, stage 1 is in place primarily to give users time to understand the format of the data and test their workflows for building models.
\n\nWe will only have a few days to incorporate the new data provided in stage 2 and predict the answers for this years tournament. So you will almost certainly want to retrain your model in stage 2 using the new data, but stage 1 allows you to test your methods so that you can do it quickly in stage 2.
", "votes": null, "replies": [] }, { "id": 765505, "author_name": "nroman", "author_url": "", "post_date": "03/06/2020 17:57:43", "content": "Basically yes, you got the idea. Stage one is mostly to play around. I have posted true labels so people can validate their models offline, without need to submit every time.
", "votes": null, "replies": [] }, { "id": 765719, "author_name": "shahules", "author_url": "", "post_date": "03/07/2020 01:55:57", "content": "Yes, that's good.But without removing the entire 2015-2019 data from the train you can use data till 2016 to train and predict for 2017 and vice versa.This is a much better approach..
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "765476": "Do I understand correctly that we need to prediction the 2015-2019 seasons? In order to avoid leakage, we should the 2015-2019 seasons removed from the training dataset and then it is worth building a model that best predicts these seasons, according data from 1985-2014 seasons. Is that right?\n\nSee my kernel: [MM NCAAM [No leaks]: LGB, XGB, LogReg](https://www.kaggle.com/vbmokin/mm-ncaam-no-leaks-lgb-xgb-logreg)\n\nAlso answers there are in the dataset [NCAA 2015-2019 labels](https://www.kaggle.com/nroman/ncaa-20152019-labels) from @nroman (really played matches of 2015-2019). \n\nOr in the post \"[How to score 0 and why you should not do this](https://www.kaggle.com/c/google-cloud-ncaa-march-madness-2020-division-1-mens-tournament/discussion/131539)\" from @nroman, there is an instruction on exactly how to read it data from data of the competition. \n\nDid I understand everything correctly?", "765500": "Yes, stage 1 is in place primarily to give users time to understand the format of the data and test their workflows for building models.\n\nWe will only have a few days to incorporate the new data provided in stage 2 and predict the answers for this years tournament. So you will almost certainly want to retrain your model in stage 2 using the new data, but stage 1 allows you to test your methods so that you can do it quickly in stage 2.", "765505": "Basically yes, you got the idea. Stage one is mostly to play around. I have posted true labels so people can validate their models offline, without need to submit every time.", "765719": "Yes, that's good.But without removing the entire 2015-2019 data from the train you can use data till 2016 to train and predict for 2017 and vice versa.This is a much better approach.." }, "source": "meta" }