{ "id": 189293, "title": "Huge changes & shout out to some legends", "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/189293", "author_name": "", "post_date": "2020-10-07T07:33:35.076544300Z", "votes": 5, "comment_count": 4, "views": 0, "content": "

Gosh!. Well I predicted my own fall fairly closely (from 65th on public LB to 589th on private LB 😬😁)

\n

But I really didn't expect the top 50 to change SO MUCH. Almost all of the current top 50 jumped by 500 to 1500 places - that's amazing. I know everyone says trust your cv score but is it really true that all of those in the top 50 now were confident that their cv scores were so solid? If so I am really impressed. Either way I have truly learnt my lesson not to chase the public LB!

\n

I am bit sad though that some people who were doing truly awesome and original work fell down the LB so much:
\n @souza - that really makes me sad - your work was inspiring!
\n@ulrich07 - so many people were using your work and you introduced (I think) a method to many people.

\n

But I want to say thanks to them and am at least glad that some of the people who shared some really useful work (not a comprehensive list - just some that helped me) got bronzes at least:
\n@allunia
\n@gunesevitan
\n@hfutybx

\n

Thanks all - looking forward to reading how you guys did it and really, really hoping the CTscans were useful after all….

", "messages": [ { "id": "1040524", "postDate": "10/07/2020 07:33:35", "content": "

Gosh!. Well I predicted my own fall fairly closely (from 65th on public LB to 589th on private LB 😬😁)

\n

But I really didn't expect the top 50 to change SO MUCH. Almost all of the current top 50 jumped by 500 to 1500 places - that's amazing. I know everyone says trust your cv score but is it really true that all of those in the top 50 now were confident that their cv scores were so solid? If so I am really impressed. Either way I have truly learnt my lesson not to chase the public LB!

\n

I am bit sad though that some people who were doing truly awesome and original work fell down the LB so much:
\n @souza - that really makes me sad - your work was inspiring!
\n@ulrich07 - so many people were using your work and you introduced (I think) a method to many people.

\n

But I want to say thanks to them and am at least glad that some of the people who shared some really useful work (not a comprehensive list - just some that helped me) got bronzes at least:
\n@allunia
\n@gunesevitan
\n@hfutybx

\n

Thanks all - looking forward to reading how you guys did it and really, really hoping the CTscans were useful after all….

", "rawMarkdown": "Gosh!. Well I predicted my own fall fairly closely (from 65th on public LB to 589th on private LB 😬😁)\n\nBut I really didn't expect the top 50 to change SO MUCH. Almost all of the current top 50 jumped by 500 to 1500 places - that's amazing. I know everyone says trust your cv score but is it really true that all of those in the top 50 now were confident that their cv scores were so solid? If so I am really impressed. Either way I have truly learnt my lesson not to chase the public LB!\n\nI am bit sad though that some people who were doing truly awesome and original work fell down the LB so much:\n @souza - that really makes me sad - your work was inspiring!\n@ulrich07 - so many people were using your work and you introduced (I think) a method to many people.\n\nBut I want to say thanks to them and am at least glad that some of the people who shared some really useful work (not a comprehensive list - just some that helped me) got bronzes at least:\n@allunia \n@gunesevitan \n@hfutybx \n\nThanks all - looking forward to reading how you guys did it and really, really hoping the CTscans were useful after all....", "votes": null }, { "id": "1040535", "postDate": "10/07/2020 07:52:02", "content": "

Yes, we never looked at out leaderbord score :)

", "rawMarkdown": "Yes, we never looked at out leaderbord score :)", "votes": null }, { "id": "1040945", "postDate": "10/07/2020 13:12:49", "content": "

I got 21 features from here to get final scores.Even though my method didn't perfect,these features were meaningful for results and helped me get bronzes.
\nI have updated notebook and glad it helped you.I also learned a lot from the competition.

", "rawMarkdown": "I got 21 features from [here](https://www.kaggle.com/hfutybx/residual-train-pytorch-osic-multiple-quantile?scriptVersionId=44169476) to get final scores.Even though my method didn't perfect,these features were meaningful for results and helped me get bronzes.\nI have updated [notebook](https://www.kaggle.com/hfutybx/osic-feature-extract-from-ct) and glad it helped you.I also learned a lot from the competition.", "votes": null }, { "id": "1040981", "postDate": "10/07/2020 13:33:53", "content": "

Hi @cascadenite ,

\n

thank you for shouting out my work. :-) I was some of the lucky ones that trusted their CV. I used @souza bayesian hierarchical linear regression model but tuned some hyperparameters on my own. ;-)

\n

I tried a lot of more models and ideas improve my result but regardless of what I tried my CV-scores hardly improved. For this reason I stayed with the bayesian linear model and I found it so elegant for this data that I wanted to see how this behaves on the private LB. Sadly the data had too less samples per patient and only one CT-scan per patient has a very high chance for overfitting badly when using it. I haven't found a good way to use it.

", "rawMarkdown": "Hi @cascadenite , \n\nthank you for shouting out my work. :-) I was some of the lucky ones that trusted their CV. I used @souza bayesian hierarchical linear regression model but tuned some hyperparameters on my own. ;-)\n\nI tried a lot of more models and ideas improve my result but regardless of what I tried my CV-scores hardly improved. For this reason I stayed with the bayesian linear model and I found it so elegant for this data that I wanted to see how this behaves on the private LB. Sadly the data had too less samples per patient and only one CT-scan per patient has a very high chance for overfitting badly when using it. I haven't found a good way to use it.", "votes": null }, { "id": "1041592", "postDate": "10/07/2020 20:36:36", "content": "

I have to protect @souza, I choose the models to be evaluate in the end. We had models in 43rd place and 100ish, but as everyone in the top, I was blind by the LB. We have learned a lot in this competition, but the main lesson is: believe in the theory. We knew that using percent was wrong.

", "rawMarkdown": "I have to protect @souza, I choose the models to be evaluate in the end. We had models in 43rd place and 100ish, but as everyone in the top, I was blind by the LB. We have learned a lot in this competition, but the main lesson is: believe in the theory. We knew that using percent was wrong.", "votes": null } ], "comments": [ { "id": 1040535, "author_name": "lukereijnen", "author_url": "", "post_date": "10/07/2020 07:52:02", "content": "

Yes, we never looked at out leaderbord score :)

", "votes": null, "replies": [] }, { "id": 1040945, "author_name": "hfutybx", "author_url": "", "post_date": "10/07/2020 13:12:49", "content": "

I got 21 features from here to get final scores.Even though my method didn't perfect,these features were meaningful for results and helped me get bronzes.
\nI have updated notebook and glad it helped you.I also learned a lot from the competition.

", "votes": null, "replies": [] }, { "id": 1040981, "author_name": "allunia", "author_url": "", "post_date": "10/07/2020 13:33:53", "content": "

Hi @cascadenite ,

\n

thank you for shouting out my work. :-) I was some of the lucky ones that trusted their CV. I used @souza bayesian hierarchical linear regression model but tuned some hyperparameters on my own. ;-)

\n

I tried a lot of more models and ideas improve my result but regardless of what I tried my CV-scores hardly improved. For this reason I stayed with the bayesian linear model and I found it so elegant for this data that I wanted to see how this behaves on the private LB. Sadly the data had too less samples per patient and only one CT-scan per patient has a very high chance for overfitting badly when using it. I haven't found a good way to use it.

", "votes": null, "replies": [] }, { "id": 1041592, "author_name": "cafalchio", "author_url": "", "post_date": "10/07/2020 20:36:36", "content": "

I have to protect @souza, I choose the models to be evaluate in the end. We had models in 43rd place and 100ish, but as everyone in the top, I was blind by the LB. We have learned a lot in this competition, but the main lesson is: believe in the theory. We knew that using percent was wrong.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1040524": "Gosh!. Well I predicted my own fall fairly closely (from 65th on public LB to 589th on private LB 😬😁)\n\nBut I really didn't expect the top 50 to change SO MUCH. Almost all of the current top 50 jumped by 500 to 1500 places - that's amazing. I know everyone says trust your cv score but is it really true that all of those in the top 50 now were confident that their cv scores were so solid? If so I am really impressed. Either way I have truly learnt my lesson not to chase the public LB!\n\nI am bit sad though that some people who were doing truly awesome and original work fell down the LB so much:\n @souza - that really makes me sad - your work was inspiring!\n@ulrich07 - so many people were using your work and you introduced (I think) a method to many people.\n\nBut I want to say thanks to them and am at least glad that some of the people who shared some really useful work (not a comprehensive list - just some that helped me) got bronzes at least:\n@allunia \n@gunesevitan \n@hfutybx \n\nThanks all - looking forward to reading how you guys did it and really, really hoping the CTscans were useful after all....", "1040535": "Yes, we never looked at out leaderbord score :)", "1040945": "I got 21 features from [here](https://www.kaggle.com/hfutybx/residual-train-pytorch-osic-multiple-quantile?scriptVersionId=44169476) to get final scores.Even though my method didn't perfect,these features were meaningful for results and helped me get bronzes.\nI have updated [notebook](https://www.kaggle.com/hfutybx/osic-feature-extract-from-ct) and glad it helped you.I also learned a lot from the competition.", "1040981": "Hi @cascadenite , \n\nthank you for shouting out my work. :-) I was some of the lucky ones that trusted their CV. I used @souza bayesian hierarchical linear regression model but tuned some hyperparameters on my own. ;-)\n\nI tried a lot of more models and ideas improve my result but regardless of what I tried my CV-scores hardly improved. For this reason I stayed with the bayesian linear model and I found it so elegant for this data that I wanted to see how this behaves on the private LB. Sadly the data had too less samples per patient and only one CT-scan per patient has a very high chance for overfitting badly when using it. I haven't found a good way to use it.", "1041592": "I have to protect @souza, I choose the models to be evaluate in the end. We had models in 43rd place and 100ish, but as everyone in the top, I was blind by the LB. We have learned a lot in this competition, but the main lesson is: believe in the theory. We knew that using percent was wrong." }, "source": "meta" }