what will be the lgb interpretation for unseen continuous valued feature. For example, let's say an LGB model was trained on a continuous feature f1 having values in the range (1,10) and now it is to be tested on a test set having values for f1 in range(11,30), then how lgb will handle these new interpretation?
", "messages": [ { "id": "393667", "postDate": "09/25/2018 18:29:46", "content": "what will be the lgb interpretation for unseen continuous valued feature. For example, let's say an LGB model was trained on a continuous feature f1 having values in the range (1,10) and now it is to be tested on a test set having values for f1 in range(11,30), then how lgb will handle these new interpretation?
", "rawMarkdown": "what will be the lgb interpretation for unseen continuous valued feature. For example, let's say an LGB model was trained on a continuous feature f1 having values in the range (1,10) and now it is to be tested on a test set having values for f1 in range(11,30), then how lgb will handle these new interpretation?", "votes": null }, { "id": "393802", "postDate": "09/25/2018 22:44:28", "content": "I think this is a pretty good explanation oh how an unseen range is handled by LGB.
\n\n\n", "rawMarkdown": "I think [this][1] is a pretty good explanation oh how an unseen range is handled by LGB.\n\n> It is not aware on the bounds of the values of the feature. All it knows is \"greater than\" or \"lower than\" to choose the cut point.\n\n [1]: https://github.com/dmlc/xgboost/issues/1581#issuecomment-249846368", "votes": null }, { "id": "393925", "postDate": "09/26/2018 04:32:19", "content": "It is not aware on the bounds of the values of the feature. All it knows is \"greater than\" or \"lower than\" to choose the cut point.
\n
Thanks for the resource. but according to that, the cut point will always be >=lower bound and <=upper bound values for that feature, now in test set, if we have our distribution for this feature such that even lowest value is >= upper bound of train set, then at the point of splitting, our test set will now never split in two parts as we will always reach in region of >= upper bound, that way, this splitting node will become useless I think. correct me, if I am wrong!!
", "rawMarkdown": "Thanks for the resource. but according to that, the cut point will always be >=lower bound and <=upper bound values for that feature, now in test set, if we have our distribution for this feature such that even lowest value is >= upper bound of train set, then at the point of splitting, our test set will now never split in two parts as we will always reach in region of >= upper bound, that way, this splitting node will become useless I think. correct me, if I am wrong!!", "votes": null }, { "id": "393962", "postDate": "09/26/2018 05:55:15", "content": "You can always make experiment and see what happens. As for competition environment, the situation with unseen data can be avoided.
", "rawMarkdown": "You can always make experiment and see what happens. As for competition environment, the situation with unseen data can be avoided.", "votes": null }, { "id": "394424", "postDate": "09/26/2018 20:02:07", "content": "A caveat: LightGBM performs bucketing for continuous features if the number of unique features is < 255. This is according to the max_bin parameter. So, I wonder what happens if you have 1, 2, ..., 254 as unique continuous features then you have new values 255, 256, 257, ..., 300 in the test set. I wonder how LightGBM adapts to revert from a binning approach to a numerical splits approach. Hm! If anyone knows please comment
As for competition environment, the situation with unseen data can be avoided
\n\nWell I don't think so!! That will not help at all.
", "rawMarkdown": "*As for competition environment, the situation with unseen data can be avoided* \n\nWell I don't think so!! That will not help at all.", "votes": null } ], "comments": [ { "id": 393802, "author_name": "nnnnick", "author_url": "", "post_date": "09/25/2018 22:44:28", "content": "I think this is a pretty good explanation oh how an unseen range is handled by LGB.
\n\n\n", "votes": null, "replies": [ { "id": 393925, "author_name": "mk9440", "author_url": "", "post_date": "09/26/2018 04:32:19", "content": "It is not aware on the bounds of the values of the feature. All it knows is \"greater than\" or \"lower than\" to choose the cut point.
\n
Thanks for the resource. but according to that, the cut point will always be >=lower bound and <=upper bound values for that feature, now in test set, if we have our distribution for this feature such that even lowest value is >= upper bound of train set, then at the point of splitting, our test set will now never split in two parts as we will always reach in region of >= upper bound, that way, this splitting node will become useless I think. correct me, if I am wrong!!
", "votes": null, "replies": [] }, { "id": 394424, "author_name": "returnofsputnik", "author_url": "", "post_date": "09/26/2018 20:02:07", "content": "A caveat: LightGBM performs bucketing for continuous features if the number of unique features is < 255. This is according to the max_bin parameter. So, I wonder what happens if you have 1, 2, ..., 254 as unique continuous features then you have new values 255, 256, 257, ..., 300 in the test set. I wonder how LightGBM adapts to revert from a binning approach to a numerical splits approach. Hm! If anyone knows please comment
You can always make experiment and see what happens. As for competition environment, the situation with unseen data can be avoided.
", "votes": null, "replies": [ { "id": 394437, "author_name": "mk9440", "author_url": "", "post_date": "09/26/2018 20:46:11", "content": "As for competition environment, the situation with unseen data can be avoided
\n\nWell I don't think so!! That will not help at all.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "393667": "what will be the lgb interpretation for unseen continuous valued feature. For example, let's say an LGB model was trained on a continuous feature f1 having values in the range (1,10) and now it is to be tested on a test set having values for f1 in range(11,30), then how lgb will handle these new interpretation?", "393802": "I think [this][1] is a pretty good explanation oh how an unseen range is handled by LGB.\n\n> It is not aware on the bounds of the values of the feature. All it knows is \"greater than\" or \"lower than\" to choose the cut point.\n\n [1]: https://github.com/dmlc/xgboost/issues/1581#issuecomment-249846368", "393925": "Thanks for the resource. but according to that, the cut point will always be >=lower bound and <=upper bound values for that feature, now in test set, if we have our distribution for this feature such that even lowest value is >= upper bound of train set, then at the point of splitting, our test set will now never split in two parts as we will always reach in region of >= upper bound, that way, this splitting node will become useless I think. correct me, if I am wrong!!", "393962": "You can always make experiment and see what happens. As for competition environment, the situation with unseen data can be avoided.", "394424": "A caveat: LightGBM performs bucketing for continuous features if the number of unique features is < 255. This is according to the `max_bin` parameter. So, I wonder what happens if you have `1, 2, ..., 254` as unique continuous features then you have new values `255, 256, 257, ..., 300` in the test set. I wonder how LightGBM adapts to revert from a binning approach to a numerical splits approach. Hm! If anyone knows please comment", "394437": "*As for competition environment, the situation with unseen data can be avoided* \n\nWell I don't think so!! That will not help at all." }, "source": "meta" }