I'm still thinking about using Percent feature instead of the first value. What are your opinions regarding this? I know it is wrong and it doesn't make sense but is there any scenario it'll work better than Baseline Percent?
", "messages": [ { "id": "1038870", "postDate": "10/06/2020 06:08:23", "content": "I'm still thinking about using Percent feature instead of the first value. What are your opinions regarding this? I know it is wrong and it doesn't make sense but is there any scenario it'll work better than Baseline Percent?
", "rawMarkdown": "I'm still thinking about using Percent feature instead of the first value. What are your opinions regarding this? I know it is wrong and it doesn't make sense but is there any scenario it'll work better than Baseline Percent?", "votes": null }, { "id": "1038922", "postDate": "10/06/2020 07:03:31", "content": "Does it actually help in CV, when out-of-fold you do not have it available and use baseline instead? That would seem like the only real test for the private LB. And then, one could still mistrust it due to not really being the same one will get on the LB.
", "rawMarkdown": "Does it actually help in CV, when out-of-fold you do not have it available and use baseline instead? That would seem like the only real test for the private LB. And then, one could still mistrust it due to not really being the same one will get on the LB.", "votes": null }, { "id": "1038988", "postDate": "10/06/2020 08:19:57", "content": "Add percent improve my score in CV and LB so i use it
", "rawMarkdown": "Add percent improve my score in CV and LB so i use it", "votes": null }, { "id": "1039087", "postDate": "10/06/2020 10:07:01", "content": "PS: I would also be worried that it messes up stacking, because even if it helps a bit in LB and a lot in CV, solutions using the feature might get too much weight in stacking. That's of course only an issue if CV is not done with resetting percent to baseline for out-of-fold predictions.
", "rawMarkdown": "PS: I would also be worried that it messes up stacking, because even if it helps a bit in LB and a lot in CV, solutions using the feature might get too much weight in stacking. That's of course only an issue if CV is not done with resetting percent to baseline for out-of-fold predictions.", "votes": null }, { "id": "1039112", "postDate": "10/06/2020 10:39:32", "content": "What I did is I kept Percent in all my training folds and I only used init_Percent in my validation folds
", "rawMarkdown": "What I did is I kept Percent in all my training folds and I only used init_Percent in my validation folds", "votes": null }, { "id": "1039151", "postDate": "10/06/2020 11:20:37", "content": "This is what I really wanna know!
", "rawMarkdown": "This is what I really wanna know!", "votes": null }, { "id": "1039392", "postDate": "10/06/2020 14:47:01", "content": "This is the only way to actually understand if it works. If it improves CV then ok to use it.
\n@seif95 does it ?
It improves CV and I have correlation between public LB and cv score.
\nI hope I will be able to do a \"remontada\" in private :p
Yes, that seems reasonable.
\nGood luck on the final results, a very big shakeup is coming :)
I think Percent was the culprit of this shake up. All of my solutions using baseline Percent are in medal range (two of them are at #1 spot).
", "rawMarkdown": "I think Percent was the culprit of this shake up. All of my solutions using baseline Percent are in medal range (two of them are at #1 spot).", "votes": null } ], "comments": [ { "id": 1038922, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "10/06/2020 07:03:31", "content": "Does it actually help in CV, when out-of-fold you do not have it available and use baseline instead? That would seem like the only real test for the private LB. And then, one could still mistrust it due to not really being the same one will get on the LB.
", "votes": null, "replies": [ { "id": 1039087, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "10/06/2020 10:07:01", "content": "PS: I would also be worried that it messes up stacking, because even if it helps a bit in LB and a lot in CV, solutions using the feature might get too much weight in stacking. That's of course only an issue if CV is not done with resetting percent to baseline for out-of-fold predictions.
", "votes": null, "replies": [] } ] }, { "id": 1038988, "author_name": "gzl0506", "author_url": "", "post_date": "10/06/2020 08:19:57", "content": "Add percent improve my score in CV and LB so i use it
", "votes": null, "replies": [] }, { "id": 1039112, "author_name": "seif95", "author_url": "", "post_date": "10/06/2020 10:39:32", "content": "What I did is I kept Percent in all my training folds and I only used init_Percent in my validation folds
", "votes": null, "replies": [ { "id": 1039392, "author_name": "watzisname", "author_url": "", "post_date": "10/06/2020 14:47:01", "content": "This is the only way to actually understand if it works. If it improves CV then ok to use it.
\n@seif95 does it ?
It improves CV and I have correlation between public LB and cv score.
\nI hope I will be able to do a \"remontada\" in private :p
Yes, that seems reasonable.
\nGood luck on the final results, a very big shakeup is coming :)
This is what I really wanna know!
", "votes": null, "replies": [] }, { "id": 1040315, "author_name": "gunesevitan", "author_url": "", "post_date": "10/07/2020 04:45:15", "content": "I think Percent was the culprit of this shake up. All of my solutions using baseline Percent are in medal range (two of them are at #1 spot).
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1038870": "I'm still thinking about using Percent feature instead of the first value. What are your opinions regarding this? I know it is wrong and it doesn't make sense but is there any scenario it'll work better than Baseline Percent?", "1038922": "Does it actually help in CV, when out-of-fold you do not have it available and use baseline instead? That would seem like the only real test for the private LB. And then, one could still mistrust it due to not really being the same one will get on the LB.", "1038988": "Add percent improve my score in CV and LB so i use it", "1039087": "PS: I would also be worried that it messes up stacking, because even if it helps a bit in LB and a lot in CV, solutions using the feature might get too much weight in stacking. That's of course only an issue if CV is not done with resetting percent to baseline for out-of-fold predictions.", "1039112": "What I did is I kept Percent in all my training folds and I only used init_Percent in my validation folds", "1039151": "This is what I really wanna know!", "1039392": "This is the only way to actually understand if it works. If it improves CV then ok to use it.\n@seif95 does it ?", "1039398": "It improves CV and I have correlation between public LB and cv score. \nI hope I will be able to do a \"remontada\" in private :p", "1039415": "Yes, that seems reasonable. \nGood luck on the final results, a very big shakeup is coming :)", "1040315": "I think Percent was the culprit of this shake up. All of my solutions using baseline Percent are in medal range (two of them are at #1 spot)." }, "source": "meta" }