My observation with top 100 rankings in Pvt LB, only 7 teams rank has decreased . Rest all of them have risen from no where in public LB. Is this very common in kaggle competitions? Any diagnosis report on why this kind of shake up in Pvt LB vs Public LB?? Any pointers on how to be stable with the Pvt LB and Pub LB??
", "messages": [ { "id": "975114", "postDate": "08/18/2020 06:46:55", "content": "My observation with top 100 rankings in Pvt LB, only 7 teams rank has decreased . Rest all of them have risen from no where in public LB. Is this very common in kaggle competitions? Any diagnosis report on why this kind of shake up in Pvt LB vs Public LB?? Any pointers on how to be stable with the Pvt LB and Pub LB??
", "rawMarkdown": "My observation with top 100 rankings in Pvt LB, only 7 teams rank has decreased . Rest all of them have risen from no where in public LB. Is this very common in kaggle competitions? Any diagnosis report on why this kind of shake up in Pvt LB vs Public LB?? Any pointers on how to be stable with the Pvt LB and Pub LB??", "votes": null }, { "id": "975128", "postDate": "08/18/2020 07:00:01", "content": "There is no correlation between public LB and the private one. I've already opened the discussion about a potential shake-up. There I've sad that will get a huge shake-up like in M5. However, I didn't expect so ridiculous results. I just dunno how it's possible when all medal owners from public LB gave them away to outsiders.
", "rawMarkdown": "There is no correlation between public LB and the private one. I've already opened the discussion about a potential shake-up. There I've sad that will get a huge shake-up like in M5. However, I didn't expect so ridiculous results. I just dunno how it's possible when all medal owners from public LB gave them away to outsiders.", "votes": null }, { "id": "975196", "postDate": "08/18/2020 07:33:57", "content": "A lot of them forked the high ranking blends like 0.9619 and performed further blends on them - btw, 9619 performed terribly on the Private Leaderboard.
\nOthers were clear overfits and were probing leaderboard. Also, the margin for error was very small here and hence the shakeup.
", "rawMarkdown": "A lot of them forked the high ranking blends like 0.9619 and performed further blends on them - btw, 9619 performed terribly on the Private Leaderboard.\n\nOthers were clear overfits and were probing leaderboard. Also, the margin for error was very small here and hence the shakeup.", "votes": null }, { "id": "975253", "postDate": "08/18/2020 08:06:24", "content": "Specific situation, it happens time after time, but it is not so common. Here we had small positive class that was too easy to overfit. One proposed to trust your CV but when Test set and Train set differ much and there is some cluster of mystery images, then maybe LB would be better than CV. In this comp we see that CV beat LB
", "rawMarkdown": "Specific situation, it happens time after time, but it is not so common. Here we had small positive class that was too easy to overfit. One proposed to trust your CV but when Test set and Train set differ much and there is some cluster of mystery images, then maybe LB would be better than CV. In this comp we see that CV beat LB", "votes": null } ], "comments": [ { "id": 975128, "author_name": "vadimtimakin", "author_url": "", "post_date": "08/18/2020 07:00:01", "content": "There is no correlation between public LB and the private one. I've already opened the discussion about a potential shake-up. There I've sad that will get a huge shake-up like in M5. However, I didn't expect so ridiculous results. I just dunno how it's possible when all medal owners from public LB gave them away to outsiders.
", "votes": null, "replies": [] }, { "id": 975196, "author_name": "vishram6", "author_url": "", "post_date": "08/18/2020 07:33:57", "content": "A lot of them forked the high ranking blends like 0.9619 and performed further blends on them - btw, 9619 performed terribly on the Private Leaderboard.
\nOthers were clear overfits and were probing leaderboard. Also, the margin for error was very small here and hence the shakeup.
", "votes": null, "replies": [] }, { "id": 975253, "author_name": "aybatov", "author_url": "", "post_date": "08/18/2020 08:06:24", "content": "Specific situation, it happens time after time, but it is not so common. Here we had small positive class that was too easy to overfit. One proposed to trust your CV but when Test set and Train set differ much and there is some cluster of mystery images, then maybe LB would be better than CV. In this comp we see that CV beat LB
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "975114": "My observation with top 100 rankings in Pvt LB, only 7 teams rank has decreased . Rest all of them have risen from no where in public LB. Is this very common in kaggle competitions? Any diagnosis report on why this kind of shake up in Pvt LB vs Public LB?? Any pointers on how to be stable with the Pvt LB and Pub LB??", "975128": "There is no correlation between public LB and the private one. I've already opened the discussion about a potential shake-up. There I've sad that will get a huge shake-up like in M5. However, I didn't expect so ridiculous results. I just dunno how it's possible when all medal owners from public LB gave them away to outsiders.", "975196": "A lot of them forked the high ranking blends like 0.9619 and performed further blends on them - btw, 9619 performed terribly on the Private Leaderboard.\n\nOthers were clear overfits and were probing leaderboard. Also, the margin for error was very small here and hence the shakeup.", "975253": "Specific situation, it happens time after time, but it is not so common. Here we had small positive class that was too easy to overfit. One proposed to trust your CV but when Test set and Train set differ much and there is some cluster of mystery images, then maybe LB would be better than CV. In this comp we see that CV beat LB" }, "source": "meta" }