When doing feature engineering one should think about whether a particular feature would be useful not only in training set but, and this is much more important, also in testing set.
\n\nHere are couple of features that don't work (will try to keep this updated):
\n\nTeam's previous game results, i.e. whether the team won previous game (1) or lost it (0). - This would not work, because in stage 2 we will predict the NCAA tourney, which is a play-off. This means that both teams will always have 1 for this feature.
Whether the game will be held at Team1 hometown (1) or Team2 hometown (2) or at a neutral territory (0). - Since all the games (or at least most of them) in NCAA are being held at a neutral territory this feature will always be 0, thus would be useless.
Feel free to share an ideas of your own.
", "messages": [ { "id": "764308", "postDate": "03/05/2020 10:22:47", "content": "When doing feature engineering one should think about whether a particular feature would be useful not only in training set but, and this is much more important, also in testing set.
\n\nHere are couple of features that don't work (will try to keep this updated):
\n\nTeam's previous game results, i.e. whether the team won previous game (1) or lost it (0). - This would not work, because in stage 2 we will predict the NCAA tourney, which is a play-off. This means that both teams will always have 1 for this feature.
Whether the game will be held at Team1 hometown (1) or Team2 hometown (2) or at a neutral territory (0). - Since all the games (or at least most of them) in NCAA are being held at a neutral territory this feature will always be 0, thus would be useless.
Feel free to share an ideas of your own.
", "rawMarkdown": "When doing feature engineering one should think about whether a particular feature would be useful not only in training set but, and this is much more important, also in testing set.\n\nHere are couple of features that don't work (will try to keep this updated):\n\n1. Team's previous game results, i.e. whether the team won previous game (1) or lost it (0). - This would not work, because in stage 2 we will predict the NCAA tourney, which is a play-off. This means that both teams will always have 1 for this feature.\n\n2. Whether the game will be held at Team1 hometown (1) or Team2 hometown (2) or at a neutral territory (0). - Since all the games (or at least most of them) in NCAA are being held at a neutral territory this feature will always be 0, thus would be useless.\n\nFeel free to share an ideas of your own.", "votes": null }, { "id": "764492", "postDate": "03/05/2020 14:20:43", "content": "Every game will be at a \"neutral\" site but it's possible for teams to still play close to home and thus have strong fan support. If someone put the time into making a notebook that built this into a feature, it would be really interesting.
\n\nAlso, to your first point, margin of victory of a previous game also applies.
", "rawMarkdown": "Every game will be at a \"neutral\" site but it's possible for teams to still play close to home and thus have strong fan support. If someone put the time into making a notebook that built this into a feature, it would be really interesting. \n\nAlso, to your first point, margin of victory of a previous game also applies.", "votes": null }, { "id": "765626", "postDate": "03/06/2020 22:17:40", "content": "Maybe. But those features mentioned above really don't work.
", "rawMarkdown": "Maybe. But those features mentioned above really don't work.", "votes": null }, { "id": "766240", "postDate": "03/07/2020 22:34:42", "content": "#2 is entirely dependent on how the user choses to label each tournament game since that label is not provided by Kaggle in the test set. Technically, the NCAA defines all of the tournament games as neutral, but as @btesch points out, there are other ways to determine whether the game location is truly \"neutral.\" So the feature could work depending on how the test labels were generated.
", "rawMarkdown": "\\#2 is entirely dependent on how the user choses to label each tournament game since that label is not provided by Kaggle in the test set. Technically, the NCAA defines all of the tournament games as neutral, but as @btesch points out, there are other ways to determine whether the game location is truly \"neutral.\" So the feature could work depending on how the test labels were generated.", "votes": null } ], "comments": [ { "id": 764492, "author_name": "btesch", "author_url": "", "post_date": "03/05/2020 14:20:43", "content": "Every game will be at a \"neutral\" site but it's possible for teams to still play close to home and thus have strong fan support. If someone put the time into making a notebook that built this into a feature, it would be really interesting.
\n\nAlso, to your first point, margin of victory of a previous game also applies.
", "votes": null, "replies": [ { "id": 765626, "author_name": "vbmokin", "author_url": "", "post_date": "03/06/2020 22:17:40", "content": "Maybe. But those features mentioned above really don't work.
", "votes": null, "replies": [] }, { "id": 766240, "author_name": "rsa013", "author_url": "", "post_date": "03/07/2020 22:34:42", "content": "#2 is entirely dependent on how the user choses to label each tournament game since that label is not provided by Kaggle in the test set. Technically, the NCAA defines all of the tournament games as neutral, but as @btesch points out, there are other ways to determine whether the game location is truly \"neutral.\" So the feature could work depending on how the test labels were generated.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "764308": "When doing feature engineering one should think about whether a particular feature would be useful not only in training set but, and this is much more important, also in testing set.\n\nHere are couple of features that don't work (will try to keep this updated):\n\n1. Team's previous game results, i.e. whether the team won previous game (1) or lost it (0). - This would not work, because in stage 2 we will predict the NCAA tourney, which is a play-off. This means that both teams will always have 1 for this feature.\n\n2. Whether the game will be held at Team1 hometown (1) or Team2 hometown (2) or at a neutral territory (0). - Since all the games (or at least most of them) in NCAA are being held at a neutral territory this feature will always be 0, thus would be useless.\n\nFeel free to share an ideas of your own.", "764492": "Every game will be at a \"neutral\" site but it's possible for teams to still play close to home and thus have strong fan support. If someone put the time into making a notebook that built this into a feature, it would be really interesting. \n\nAlso, to your first point, margin of victory of a previous game also applies.", "765626": "Maybe. But those features mentioned above really don't work.", "766240": "\\#2 is entirely dependent on how the user choses to label each tournament game since that label is not provided by Kaggle in the test set. Technically, the NCAA defines all of the tournament games as neutral, but as @btesch points out, there are other ways to determine whether the game location is truly \"neutral.\" So the feature could work depending on how the test labels were generated." }, "source": "meta" }