{ "id": 220607, "title": "[0.901 PB solution] From gold to abyss or history on UnLucky leaf", "url": "/competitions/cassava-leaf-disease-classification/writeups/lucky-leaf-0-901-pb-solution-from-gold-to-abyss-or", "author_name": "", "post_date": "2021-02-19T00:44:13.025150Z", "votes": 10, "comment_count": 3, "views": 0, "content": "

Hi everyone,
\nWant share with you lesson that I learned from this competition:
\nUse only simple ensembling technics when your data is noisy

\n

This simple lesson cost a lot of places on PB for the Lucky leaf team.
\n\"\"

\n

Shortly, our 0.901 PB solution is actually the same as the selected 0.889 one (here I described it in detail), but we used only a linear combination of ResNeXt, EfficientNet, and ViT instead of CatBoost on the second layer. So, I suppose, we have a good example of KISS principle in DS.

", "messages": [ { "id": "1209587", "postDate": "02/19/2021 00:44:13", "content": "

Hi everyone,
\nWant share with you lesson that I learned from this competition:
\nUse only simple ensembling technics when your data is noisy

\n

This simple lesson cost a lot of places on PB for the Lucky leaf team.
\n\"\"

\n

Shortly, our 0.901 PB solution is actually the same as the selected 0.889 one (here I described it in detail), but we used only a linear combination of ResNeXt, EfficientNet, and ViT instead of CatBoost on the second layer. So, I suppose, we have a good example of KISS principle in DS.

", "rawMarkdown": "Hi everyone,\nWant share with you lesson that I learned from this competition:\n**Use only simple ensembling technics when your data is noisy**\n\nThis simple lesson cost a lot of places on PB for the Lucky leaf team.\n![](https://i.postimg.cc/8PYGnTvv/Screenshot-2021-02-19-at-2-33-01-AM.png)\n\n\nShortly, our 0.901 PB solution is actually the same as the selected 0.889 one ([here I described it in detail](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220588)), but we used only a linear combination of ResNeXt, EfficientNet, and ViT instead of CatBoost on the second layer. So, I suppose, we have a good example of KISS principle in DS.", "votes": null }, { "id": "1209634", "postDate": "02/19/2021 01:30:06", "content": "

Thanks for sharing, is there any intuition behind why \"Use only simple ensembling technics when your data is noisy\"?

", "rawMarkdown": "Thanks for sharing, is there any intuition behind why \"Use only simple ensembling technics when your data is noisy\"?", "votes": null }, { "id": "1210717", "postDate": "02/19/2021 16:18:26", "content": "

Intuition is the following - ensembling technics with complex decision boundaries (like Catboost that we used) are fitted to deal with the train noise, but not the test one that probably has a different source (e.g. collected from a different location and annotated by other annotators). On the other hand simple ensembling (like weighed average) technics more robust to such differences in noise distributions, because their boundaries are simple.

", "rawMarkdown": "Intuition is the following - ensembling technics with complex decision boundaries (like Catboost that we used) are fitted to deal with the train noise, but not the test one that probably has a different source (e.g. collected from a different location and annotated by other annotators). On the other hand simple ensembling (like weighed average) technics more robust to such differences in noise distributions, because their boundaries are simple.", "votes": null }, { "id": "1211897", "postDate": "02/20/2021 16:35:27", "content": "

Understood, thanks

", "rawMarkdown": "Understood, thanks", "votes": null } ], "comments": [ { "id": 1209634, "author_name": "khursani8", "author_url": "", "post_date": "02/19/2021 01:30:06", "content": "

Thanks for sharing, is there any intuition behind why \"Use only simple ensembling technics when your data is noisy\"?

", "votes": null, "replies": [ { "id": 1210717, "author_name": "alexkirnas", "author_url": "", "post_date": "02/19/2021 16:18:26", "content": "

Intuition is the following - ensembling technics with complex decision boundaries (like Catboost that we used) are fitted to deal with the train noise, but not the test one that probably has a different source (e.g. collected from a different location and annotated by other annotators). On the other hand simple ensembling (like weighed average) technics more robust to such differences in noise distributions, because their boundaries are simple.

", "votes": null, "replies": [] }, { "id": 1211897, "author_name": "khursani8", "author_url": "", "post_date": "02/20/2021 16:35:27", "content": "

Understood, thanks

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1209587": "Hi everyone,\nWant share with you lesson that I learned from this competition:\n**Use only simple ensembling technics when your data is noisy**\n\nThis simple lesson cost a lot of places on PB for the Lucky leaf team.\n![](https://i.postimg.cc/8PYGnTvv/Screenshot-2021-02-19-at-2-33-01-AM.png)\n\n\nShortly, our 0.901 PB solution is actually the same as the selected 0.889 one ([here I described it in detail](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220588)), but we used only a linear combination of ResNeXt, EfficientNet, and ViT instead of CatBoost on the second layer. So, I suppose, we have a good example of KISS principle in DS.", "1209634": "Thanks for sharing, is there any intuition behind why \"Use only simple ensembling technics when your data is noisy\"?", "1210717": "Intuition is the following - ensembling technics with complex decision boundaries (like Catboost that we used) are fitted to deal with the train noise, but not the test one that probably has a different source (e.g. collected from a different location and annotated by other annotators). On the other hand simple ensembling (like weighed average) technics more robust to such differences in noise distributions, because their boundaries are simple.", "1211897": "Understood, thanks" }, "source": "meta" }