![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1325831%2F69df58540a3db5de1f336b4f5a129f5e%2F1.jpg?generation=1563122765686404&alt=media\")
Hi, everyone!\nI'm not sure about using this technique in ML competition, but does it work to approximate error curve with any other curve to reduce the error? As I see there is a big errors on the extreme parts of graphs that can be fixed by simple post processing. Are there any visible problems in the method? it seems similar to target encoding where we use distribution of target but the other way.
\n\n
Hi, everyone!\nI'm not sure about using this technique in ML competition, but does it work to approximate error curve with any other curve to reduce the error? As I see there is a big errors on the extreme parts of graphs that can be fixed by simple post processing. Are there any visible problems in the method? it seems similar to target encoding where we use distribution of target but the other way.
\n\n
Forgot to add that the question is about post-processing of lightgbm output on validation
", "rawMarkdown": "Forgot to add that the question is about post-processing of lightgbm output on validation", "votes": null }, { "id": "578209", "postDate": "07/17/2019 13:19:31", "content": "You can put the validation set prediction in a 2nd stage model as a feature along with your original features. Then your the lightgbm can decide how to use it in training. I did this, but i didn't plot it to see how the predictions look before and after. I believe it improved my CV, though I did make other changes at the same time.
\n\nMy concern with a manual adjustment to your prediction based on the validation set plot is that maybe it won't generalize well, could be a form of overfitting I'm thinking. But it would be really neat to see what happens to your submission! Try it!
", "rawMarkdown": "You can put the validation set prediction in a 2nd stage model as a feature along with your original features. Then your the lightgbm can decide how to use it in training. I did this, but i didn't plot it to see how the predictions look before and after. I believe it improved my CV, though I did make other changes at the same time.\n\nMy concern with a manual adjustment to your prediction based on the validation set plot is that maybe it won't generalize well, could be a form of overfitting I'm thinking. But it would be really neat to see what happens to your submission! Try it!", "votes": null }, { "id": "578555", "postDate": "07/17/2019 21:07:28", "content": "Is it called oof feature?
", "rawMarkdown": "Is it called oof feature?", "votes": null }, { "id": "579065", "postDate": "07/18/2019 12:54:05", "content": "I'm not sure what it's called. But yeah that description is valid, a feature that is a prediction on OOF data. Maybe could be a form of blending, because you could include more than one OOF prediction in your second stage model.
", "rawMarkdown": "I'm not sure what it's called. But yeah that description is valid, a feature that is a prediction on OOF data. Maybe could be a form of blending, because you could include more than one OOF prediction in your second stage model.", "votes": null }, { "id": "579109", "postDate": "07/18/2019 14:01:30", "content": "it is
", "rawMarkdown": "it is", "votes": null } ], "comments": [ { "id": 574898, "author_name": "kirill702b", "author_url": "", "post_date": "07/14/2019 16:52:42", "content": "Forgot to add that the question is about post-processing of lightgbm output on validation
", "votes": null, "replies": [ { "id": 578209, "author_name": "andyatkinson", "author_url": "", "post_date": "07/17/2019 13:19:31", "content": "You can put the validation set prediction in a 2nd stage model as a feature along with your original features. Then your the lightgbm can decide how to use it in training. I did this, but i didn't plot it to see how the predictions look before and after. I believe it improved my CV, though I did make other changes at the same time.
\n\nMy concern with a manual adjustment to your prediction based on the validation set plot is that maybe it won't generalize well, could be a form of overfitting I'm thinking. But it would be really neat to see what happens to your submission! Try it!
", "votes": null, "replies": [] }, { "id": 578555, "author_name": "kirill702b", "author_url": "", "post_date": "07/17/2019 21:07:28", "content": "Is it called oof feature?
", "votes": null, "replies": [] }, { "id": 579065, "author_name": "andyatkinson", "author_url": "", "post_date": "07/18/2019 12:54:05", "content": "I'm not sure what it's called. But yeah that description is valid, a feature that is a prediction on OOF data. Maybe could be a form of blending, because you could include more than one OOF prediction in your second stage model.
", "votes": null, "replies": [] }, { "id": 579109, "author_name": "rftexas", "author_url": "", "post_date": "07/18/2019 14:01:30", "content": "it is
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "574896": "Hi, everyone!\nI'm not sure about using this technique in ML competition, but does it work to approximate error curve with any other curve to reduce the error? As I see there is a big errors on the extreme parts of graphs that can be fixed by simple post processing. Are there any visible problems in the method? it seems similar to target encoding where we use distribution of target but the other way.\n\n", "574898": "Forgot to add that the question is about post-processing of lightgbm output on validation", "578209": "You can put the validation set prediction in a 2nd stage model as a feature along with your original features. Then your the lightgbm can decide how to use it in training. I did this, but i didn't plot it to see how the predictions look before and after. I believe it improved my CV, though I did make other changes at the same time.\n\nMy concern with a manual adjustment to your prediction based on the validation set plot is that maybe it won't generalize well, could be a form of overfitting I'm thinking. But it would be really neat to see what happens to your submission! Try it!", "578555": "Is it called oof feature?", "579065": "I'm not sure what it's called. But yeah that description is valid, a feature that is a prediction on OOF data. Maybe could be a form of blending, because you could include more than one OOF prediction in your second stage model.", "579109": "it is" }, "source": "meta" }