{ "id": 189355, "title": "Another first place solution and shake up explanation", "url": "/competitions/osic-pulmonary-fibrosis-progression/writeups/fibroblaster-another-first-place-solution-and-shak", "author_name": "", "post_date": "2020-10-07T10:51:40.337548800Z", "votes": 17, "comment_count": 2, "views": 0, "content": "

These were our best private leaderboard submissions. There was no way to know they were going to win because they were my test submissions.

\n

\"i\"

\n

The best score is output of a quantile regression model trained with 25 epochs. Another thing to consider is, those models were using baseline Percent instead Percent feature. . All of the public notebooks were using 800 epochs which caused lots of overfitting on training data. Using Percent feature instead of baseline Percent caused models to output non-linear curves. I think those two things caused the shake up.

\n

I plotted oof FVC curves in those versions. You can see what kind of predictions did well on private leaderboard.
\nhttps://www.kaggle.com/gunesevitan/osic-pulmonary-fibrosis-progression-3-cv-3-model?scriptVersionId=39886853

", "messages": [ { "id": "1040789", "postDate": "10/07/2020 10:51:40", "content": "

These were our best private leaderboard submissions. There was no way to know they were going to win because they were my test submissions.

\n

\"i\"

\n

The best score is output of a quantile regression model trained with 25 epochs. Another thing to consider is, those models were using baseline Percent instead Percent feature. . All of the public notebooks were using 800 epochs which caused lots of overfitting on training data. Using Percent feature instead of baseline Percent caused models to output non-linear curves. I think those two things caused the shake up.

\n

I plotted oof FVC curves in those versions. You can see what kind of predictions did well on private leaderboard.
\nhttps://www.kaggle.com/gunesevitan/osic-pulmonary-fibrosis-progression-3-cv-3-model?scriptVersionId=39886853

", "rawMarkdown": "These were our best private leaderboard submissions. There was no way to know they were going to win because they were my test submissions.\n\n![i](https://i.ibb.co/khVpqpb/subs.jpg)\n\nThe best score is output of a quantile regression model trained with 25 epochs. Another thing to consider is, those models were using baseline Percent instead Percent feature. . All of the public notebooks were using 800 epochs which caused lots of overfitting on training data. Using Percent feature instead of baseline Percent caused models to output non-linear curves. I think those two things caused the shake up.\n\nI plotted oof FVC curves in those versions. You can see what kind of predictions did well on private leaderboard.\nhttps://www.kaggle.com/gunesevitan/osic-pulmonary-fibrosis-progression-3-cv-3-model?scriptVersionId=39886853", "votes": null }, { "id": "1040792", "postDate": "10/07/2020 10:54:36", "content": "

@gunesevitan

\n

just imagine what would we achieve with this linear one plus the image features we have =(

", "rawMarkdown": "gunesevitan \n\njust imagine what would we achieve with this linear one plus the image features we have =(", "votes": null }, { "id": "1042293", "postDate": "10/08/2020 06:58:20", "content": "

@gunesevitan ,@sandorkonya
\nyou guys are true winners.. yes it was extremely difficult to make high scoring private kernels based on public score.
\nTake away is in competition of this sort if you think your approach is robust sound,always select that submission as one of the two.

", "rawMarkdown": "gunesevitan ,@sandorkonya \nyou guys are true winners.. yes it was extremely difficult to make high scoring private kernels based on public score. \nTake away is in competition of this sort if you think your approach is robust sound,always select that submission as one of the two.", "votes": null } ], "comments": [ { "id": 1040792, "author_name": "sandorkonya", "author_url": "", "post_date": "10/07/2020 10:54:36", "content": "

@gunesevitan

\n

just imagine what would we achieve with this linear one plus the image features we have =(

", "votes": null, "replies": [] }, { "id": 1042293, "author_name": "jaideepvalani", "author_url": "", "post_date": "10/08/2020 06:58:20", "content": "

@gunesevitan ,@sandorkonya
\nyou guys are true winners.. yes it was extremely difficult to make high scoring private kernels based on public score.
\nTake away is in competition of this sort if you think your approach is robust sound,always select that submission as one of the two.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1040789": "These were our best private leaderboard submissions. There was no way to know they were going to win because they were my test submissions.\n\n![i](https://i.ibb.co/khVpqpb/subs.jpg)\n\nThe best score is output of a quantile regression model trained with 25 epochs. Another thing to consider is, those models were using baseline Percent instead Percent feature. . All of the public notebooks were using 800 epochs which caused lots of overfitting on training data. Using Percent feature instead of baseline Percent caused models to output non-linear curves. I think those two things caused the shake up.\n\nI plotted oof FVC curves in those versions. You can see what kind of predictions did well on private leaderboard.\nhttps://www.kaggle.com/gunesevitan/osic-pulmonary-fibrosis-progression-3-cv-3-model?scriptVersionId=39886853", "1040792": "gunesevitan \n\njust imagine what would we achieve with this linear one plus the image features we have =(", "1042293": "gunesevitan ,@sandorkonya \nyou guys are true winners.. yes it was extremely difficult to make high scoring private kernels based on public score. \nTake away is in competition of this sort if you think your approach is robust sound,always select that submission as one of the two." }, "source": "meta" }