I tried to make an automatic FE using the Featuretools library method and the Selection FE (SelectFromModel): https://www.kaggle.com/vbmokin/ashrae-automatic-fe-featuretools-selection-fe \nI used 3 ways for post-processing of the Featuretools rezults for part of train data:\n- Feature selection with the Pearson correlation\n- Feature selection with the SelectFromModel and LinearSVR\n- Feature selection with the SelectFromModel and RandomForestRegressor\nAfter that, I chose the common features in all three sets and just generally all of them.
\n\nAny other ideas for their automatic selection and which of these features are worth leaving?
", "messages": [ { "id": "661132", "postDate": "10/30/2019 00:27:19", "content": "I tried to make an automatic FE using the Featuretools library method and the Selection FE (SelectFromModel): https://www.kaggle.com/vbmokin/ashrae-automatic-fe-featuretools-selection-fe \nI used 3 ways for post-processing of the Featuretools rezults for part of train data:\n- Feature selection with the Pearson correlation\n- Feature selection with the SelectFromModel and LinearSVR\n- Feature selection with the SelectFromModel and RandomForestRegressor\nAfter that, I chose the common features in all three sets and just generally all of them.
\n\nAny other ideas for their automatic selection and which of these features are worth leaving?
", "rawMarkdown": "I tried to make an automatic FE using the Featuretools library method and the Selection FE (SelectFromModel): https://www.kaggle.com/vbmokin/ashrae-automatic-fe-featuretools-selection-fe \nI used 3 ways for post-processing of the Featuretools rezults for part of train data:\n- Feature selection with the Pearson correlation\n- Feature selection with the SelectFromModel and LinearSVR\n- Feature selection with the SelectFromModel and RandomForestRegressor\nAfter that, I chose the common features in all three sets and just generally all of them.\n\nAny other ideas for their automatic selection and which of these features are worth leaving?", "votes": null }, { "id": "662276", "postDate": "10/31/2019 11:14:55", "content": "I would suggest that the Pearson Correlation won't be effective for categorical -- continuous variable relationship , it would be better to work with creme's V or mutual info score in such cases .
", "rawMarkdown": "I would suggest that the Pearson Correlation won't be effective for categorical -- continuous variable relationship , it would be better to work with creme's V or mutual info score in such cases .", "votes": null }, { "id": "665280", "postDate": "11/04/2019 20:53:32", "content": "Before finding the correlation, I encoded categorical features, but I agree with you, @avikrams, the Pearson correlation gave a bad filter. Subsequently, I separately made feature selection on a separate statistical characterictics - MEDIAN and SKEW. The result came out interesting - see ASHRAE - Automatic FE&FC
", "rawMarkdown": "Before finding the correlation, I encoded categorical features, but I agree with you, @avikrams, the Pearson correlation gave a bad filter. Subsequently, I separately made feature selection on a separate statistical characterictics - MEDIAN and SKEW. The result came out interesting - see [ASHRAE - Automatic FE&FC](https://www.kaggle.com/vbmokin/ashrae-automatic-fe-fc)", "votes": null }, { "id": "665307", "postDate": "11/04/2019 21:31:12", "content": "i always found featuretools pretty hard to use and not that useful. So far I have never found any good automatic fe tools.
", "rawMarkdown": "i always found featuretools pretty hard to use and not that useful. So far I have never found any good automatic fe tools.", "votes": null }, { "id": "667310", "postDate": "11/07/2019 05:21:52", "content": "Interesting share , I made something very similar too.
", "rawMarkdown": "Interesting share , I made something very similar too.", "votes": null } ], "comments": [ { "id": 662276, "author_name": "avikrams", "author_url": "", "post_date": "10/31/2019 11:14:55", "content": "I would suggest that the Pearson Correlation won't be effective for categorical -- continuous variable relationship , it would be better to work with creme's V or mutual info score in such cases .
", "votes": null, "replies": [] }, { "id": 665280, "author_name": "vbmokin", "author_url": "", "post_date": "11/04/2019 20:53:32", "content": "Before finding the correlation, I encoded categorical features, but I agree with you, @avikrams, the Pearson correlation gave a bad filter. Subsequently, I separately made feature selection on a separate statistical characterictics - MEDIAN and SKEW. The result came out interesting - see ASHRAE - Automatic FE&FC
", "votes": null, "replies": [ { "id": 667310, "author_name": "avikrams", "author_url": "", "post_date": "11/07/2019 05:21:52", "content": "Interesting share , I made something very similar too.
", "votes": null, "replies": [] } ] }, { "id": 665307, "author_name": "lewington", "author_url": "", "post_date": "11/04/2019 21:31:12", "content": "i always found featuretools pretty hard to use and not that useful. So far I have never found any good automatic fe tools.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "661132": "I tried to make an automatic FE using the Featuretools library method and the Selection FE (SelectFromModel): https://www.kaggle.com/vbmokin/ashrae-automatic-fe-featuretools-selection-fe \nI used 3 ways for post-processing of the Featuretools rezults for part of train data:\n- Feature selection with the Pearson correlation\n- Feature selection with the SelectFromModel and LinearSVR\n- Feature selection with the SelectFromModel and RandomForestRegressor\nAfter that, I chose the common features in all three sets and just generally all of them.\n\nAny other ideas for their automatic selection and which of these features are worth leaving?", "662276": "I would suggest that the Pearson Correlation won't be effective for categorical -- continuous variable relationship , it would be better to work with creme's V or mutual info score in such cases .", "665280": "Before finding the correlation, I encoded categorical features, but I agree with you, @avikrams, the Pearson correlation gave a bad filter. Subsequently, I separately made feature selection on a separate statistical characterictics - MEDIAN and SKEW. The result came out interesting - see [ASHRAE - Automatic FE&FC](https://www.kaggle.com/vbmokin/ashrae-automatic-fe-fc)", "665307": "i always found featuretools pretty hard to use and not that useful. So far I have never found any good automatic fe tools.", "667310": "Interesting share , I made something very similar too." }, "source": "meta" }