![\"asdasd\"](\"https://i.imgur.com/hSU8d4p.png\")
Hi all, i would like to ask for some help, im kinda new to boosting algorithms. According to what i've seen in some others topics, this feature should be a good one,
\n\n
but for whatever reason, lgbm keeps putting it at the bottom of the feature importance. I'm pretty sure the other features can't be better, they are just averages,std and quantiles, and indeed they plot looks worse than this feature. Could be something related to the parameters? I would appreciate any help in this regard.
\n\nEDIT: I tried setting the importance by gain and the result it's the same, this feature appear at the bottom
", "messages": [ { "id": "536046", "postDate": "05/23/2019 21:24:55", "content": "Hi all, i would like to ask for some help, im kinda new to boosting algorithms. According to what i've seen in some others topics, this feature should be a good one,
\n\n
but for whatever reason, lgbm keeps putting it at the bottom of the feature importance. I'm pretty sure the other features can't be better, they are just averages,std and quantiles, and indeed they plot looks worse than this feature. Could be something related to the parameters? I would appreciate any help in this regard.
\n\nEDIT: I tried setting the importance by gain and the result it's the same, this feature appear at the bottom
", "rawMarkdown": "Hi all, i would like to ask for some help, im kinda new to boosting algorithms. According to what i've seen in some others topics, this feature should be a good one,\n\n\n\n but for whatever reason, lgbm keeps putting it at the bottom of the feature importance. I'm pretty sure the other features can't be better, they are just averages,std and quantiles, and indeed they plot looks worse than this feature. Could be something related to the parameters? I would appreciate any help in this regard.\n\nEDIT: I tried setting the importance by gain and the result it's the same, this feature appear at the bottom", "votes": null }, { "id": "536049", "postDate": "05/23/2019 21:29:25", "content": "Have you tried looking at gain importance? Instead of split importance.
(LightGBM) importance_type (string, optional (default=”split”)) — How the importance is calculated. If “split”, result contains numbers of times the feature is used in a model. If “gain”, result contains total gains of splits which use the feature.
", "rawMarkdown": "(LightGBM) importance_type (string, optional (default=”split”)) — How the importance is calculated. If “split”, result contains numbers of times the feature is used in a model. If “gain”, result contains total gains of splits which use the feature.", "votes": null }, { "id": "536055", "postDate": "05/23/2019 21:34:29", "content": "No, but i will try it now, i'm reading about it. Thanks :)
", "rawMarkdown": "No, but i will try it now, i'm reading about it. Thanks :)", "votes": null }, { "id": "536067", "postDate": "05/23/2019 22:01:18", "content": "The problem with this features is that it has the same value (close to zero) for ttf =15 and ttf =6 at the beginning of an EQ. So it's hard for the model to find what is the actual ttf to predict... And at the end of the EQ close to the event it seems very noisy so it's not very helpful.
", "rawMarkdown": "The problem with this features is that it has the same value (close to zero) for ttf =15 and ttf =6 at the beginning of an EQ. So it's hard for the model to find what is the actual ttf to predict... And at the end of the EQ close to the event it seems very noisy so it's not very helpful.", "votes": null }, { "id": "536068", "postDate": "05/23/2019 22:02:52", "content": "what importance fo you use, gain?
", "rawMarkdown": "what importance fo you use, gain?", "votes": null }, { "id": "536071", "postDate": "05/23/2019 22:14:52", "content": "Both now, and i get the same result. I'm pretty sure it should be ranked higher, the plots of some other features are awful and are ranked higher. Look at this one
\n\n", "rawMarkdown": "Both now, and i get the same result. I'm pretty sure it should be ranked higher, the plots of some other features are awful and are ranked higher. Look at this one\n\nhttps://imgur.com/1ph6cDC", "votes": null }, { "id": "536189", "postDate": "05/24/2019 04:42:29", "content": "You might have another feature which carries the same information but with less noise. The importance of a feature depends on the interactions with the other features too
", "rawMarkdown": "You might have another feature which carries the same information but with less noise. The importance of a feature depends on the interactions with the other features too", "votes": null }, { "id": "536294", "postDate": "05/24/2019 07:55:41", "content": "Your feature sees mini quakes as the real quakes. That maybe why it is not of high importance.
", "rawMarkdown": "Your feature sees mini quakes as the real quakes. That maybe why it is not of high importance.", "votes": null }, { "id": "536353", "postDate": "05/24/2019 09:54:28", "content": "Features that overfit training data have a high importance by definition. Unfortunately.
", "rawMarkdown": "Features that overfit training data have a high importance by definition. Unfortunately.", "votes": null } ], "comments": [ { "id": 536049, "author_name": "returnofsputnik", "author_url": "", "post_date": "05/23/2019 21:29:25", "content": "Have you tried looking at gain importance? Instead of split importance.
No, but i will try it now, i'm reading about it. Thanks :)
", "votes": null, "replies": [] } ] }, { "id": 536050, "author_name": "izmaylov", "author_url": "", "post_date": "05/23/2019 21:29:42", "content": "(LightGBM) importance_type (string, optional (default=”split”)) — How the importance is calculated. If “split”, result contains numbers of times the feature is used in a model. If “gain”, result contains total gains of splits which use the feature.
", "votes": null, "replies": [] }, { "id": 536067, "author_name": "areveillon", "author_url": "", "post_date": "05/23/2019 22:01:18", "content": "The problem with this features is that it has the same value (close to zero) for ttf =15 and ttf =6 at the beginning of an EQ. So it's hard for the model to find what is the actual ttf to predict... And at the end of the EQ close to the event it seems very noisy so it's not very helpful.
", "votes": null, "replies": [] }, { "id": 536068, "author_name": "cpmpml", "author_url": "", "post_date": "05/23/2019 22:02:52", "content": "what importance fo you use, gain?
", "votes": null, "replies": [ { "id": 536071, "author_name": "senoratiramisu", "author_url": "", "post_date": "05/23/2019 22:14:52", "content": "Both now, and i get the same result. I'm pretty sure it should be ranked higher, the plots of some other features are awful and are ranked higher. Look at this one
\n\n", "votes": null, "replies": [] }, { "id": 536353, "author_name": "cpmpml", "author_url": "", "post_date": "05/24/2019 09:54:28", "content": "Features that overfit training data have a high importance by definition. Unfortunately.
", "votes": null, "replies": [] } ] }, { "id": 536189, "author_name": "stecasasso", "author_url": "", "post_date": "05/24/2019 04:42:29", "content": "You might have another feature which carries the same information but with less noise. The importance of a feature depends on the interactions with the other features too
", "votes": null, "replies": [] }, { "id": 536294, "author_name": "cpmpml", "author_url": "", "post_date": "05/24/2019 07:55:41", "content": "Your feature sees mini quakes as the real quakes. That maybe why it is not of high importance.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "536046": "Hi all, i would like to ask for some help, im kinda new to boosting algorithms. According to what i've seen in some others topics, this feature should be a good one,\n\n\n\n but for whatever reason, lgbm keeps putting it at the bottom of the feature importance. I'm pretty sure the other features can't be better, they are just averages,std and quantiles, and indeed they plot looks worse than this feature. Could be something related to the parameters? I would appreciate any help in this regard.\n\nEDIT: I tried setting the importance by gain and the result it's the same, this feature appear at the bottom", "536049": "Have you tried looking at `gain` importance? Instead of `split` importance.", "536050": "(LightGBM) importance_type (string, optional (default=”split”)) — How the importance is calculated. If “split”, result contains numbers of times the feature is used in a model. If “gain”, result contains total gains of splits which use the feature.", "536055": "No, but i will try it now, i'm reading about it. Thanks :)", "536067": "The problem with this features is that it has the same value (close to zero) for ttf =15 and ttf =6 at the beginning of an EQ. So it's hard for the model to find what is the actual ttf to predict... And at the end of the EQ close to the event it seems very noisy so it's not very helpful.", "536068": "what importance fo you use, gain?", "536071": "Both now, and i get the same result. I'm pretty sure it should be ranked higher, the plots of some other features are awful and are ranked higher. Look at this one\n\nhttps://imgur.com/1ph6cDC", "536189": "You might have another feature which carries the same information but with less noise. The importance of a feature depends on the interactions with the other features too", "536294": "Your feature sees mini quakes as the real quakes. That maybe why it is not of high importance.", "536353": "Features that overfit training data have a high importance by definition. Unfortunately." }, "source": "meta" }