Im working with interpolation for the the variables that have fewer missing data, and that seem to work well.\nany one have other ideas ?
", "messages": [ { "id": "657249", "postDate": "10/25/2019 02:00:00", "content": "Im working with interpolation for the the variables that have fewer missing data, and that seem to work well.\nany one have other ideas ?
", "rawMarkdown": "Im working with interpolation for the the variables that have fewer missing data, and that seem to work well.\nany one have other ideas ?", "votes": null }, { "id": "657272", "postDate": "10/25/2019 02:29:35", "content": "Do you mean missing features or missing target (or both?) When I model with gradient boosting, I typically just encode null feature values as something extreme like -99999 and let the model figure out for itself how to handle nullness - this usually seems to work as well as or better than manual imputation.
\n\nFor a non tree-based model (e.g. neural net), you'd probably want to impute very carefully, since the model is probably very scale sensitive. Using column means/medians is an ok starting point, but using other feature information to increase precision (e.g. fill with mean of category groups) is often better in my experience. If you think the feature is critical to the model, it may even be worth it to take the time to build a dedicated model to predict that feature and then use this model for imputation.
", "rawMarkdown": "Do you mean missing features or missing target (or both?) When I model with gradient boosting, I typically just encode null feature values as something extreme like -99999 and let the model figure out for itself how to handle nullness - this usually seems to work as well as or better than manual imputation. \n\nFor a non tree-based model (e.g. neural net), you'd probably want to impute very carefully, since the model is probably very scale sensitive. Using column means/medians is an ok starting point, but using other feature information to increase precision (e.g. fill with mean of category groups) is often better in my experience. If you think the feature is critical to the model, it may even be worth it to take the time to build a dedicated model to predict that feature and then use this model for imputation.", "votes": null }, { "id": "658155", "postDate": "10/25/2019 18:01:52", "content": "I agree with you.\nI'm talking about features that have a few missing values like air_temperature\nim using random forest for benchmarking and feature selection with feature importance and partial dependance.
", "rawMarkdown": "I agree with you.\nI'm talking about features that have a few missing values like air_temperature\nim using random forest for benchmarking and feature selection with feature importance and partial dependance.", "votes": null }, { "id": "658190", "postDate": "10/25/2019 18:51:50", "content": "I haven't tried it yet, but filling the weather features can be done with a rolling window. I'll try to fill them with 2/3/4 hour window means and update my post here tomorrow.
", "rawMarkdown": "I haven't tried it yet, but filling the weather features can be done with a rolling window. I'll try to fill them with 2/3/4 hour window means and update my post here tomorrow.", "votes": null }, { "id": "658255", "postDate": "10/25/2019 20:55:07", "content": "great idea, thanks for sharing
", "rawMarkdown": "great idea, thanks for sharing", "votes": null } ], "comments": [ { "id": 657272, "author_name": "aquatic", "author_url": "", "post_date": "10/25/2019 02:29:35", "content": "Do you mean missing features or missing target (or both?) When I model with gradient boosting, I typically just encode null feature values as something extreme like -99999 and let the model figure out for itself how to handle nullness - this usually seems to work as well as or better than manual imputation.
\n\nFor a non tree-based model (e.g. neural net), you'd probably want to impute very carefully, since the model is probably very scale sensitive. Using column means/medians is an ok starting point, but using other feature information to increase precision (e.g. fill with mean of category groups) is often better in my experience. If you think the feature is critical to the model, it may even be worth it to take the time to build a dedicated model to predict that feature and then use this model for imputation.
", "votes": null, "replies": [] }, { "id": 658155, "author_name": "rashidhaffadi", "author_url": "", "post_date": "10/25/2019 18:01:52", "content": "I agree with you.\nI'm talking about features that have a few missing values like air_temperature\nim using random forest for benchmarking and feature selection with feature importance and partial dependance.
", "votes": null, "replies": [] }, { "id": 658190, "author_name": "gunesevitan", "author_url": "", "post_date": "10/25/2019 18:51:50", "content": "I haven't tried it yet, but filling the weather features can be done with a rolling window. I'll try to fill them with 2/3/4 hour window means and update my post here tomorrow.
", "votes": null, "replies": [ { "id": 658255, "author_name": "rashidhaffadi", "author_url": "", "post_date": "10/25/2019 20:55:07", "content": "great idea, thanks for sharing
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "657249": "Im working with interpolation for the the variables that have fewer missing data, and that seem to work well.\nany one have other ideas ?", "657272": "Do you mean missing features or missing target (or both?) When I model with gradient boosting, I typically just encode null feature values as something extreme like -99999 and let the model figure out for itself how to handle nullness - this usually seems to work as well as or better than manual imputation. \n\nFor a non tree-based model (e.g. neural net), you'd probably want to impute very carefully, since the model is probably very scale sensitive. Using column means/medians is an ok starting point, but using other feature information to increase precision (e.g. fill with mean of category groups) is often better in my experience. If you think the feature is critical to the model, it may even be worth it to take the time to build a dedicated model to predict that feature and then use this model for imputation.", "658155": "I agree with you.\nI'm talking about features that have a few missing values like air_temperature\nim using random forest for benchmarking and feature selection with feature importance and partial dependance.", "658190": "I haven't tried it yet, but filling the weather features can be done with a rolling window. I'll try to fill them with 2/3/4 hour window means and update my post here tomorrow.", "658255": "great idea, thanks for sharing" }, "source": "meta" }