I have published my first kernel in this competition on the day it started and it had quite a good score. But quite soon we could see scores lower than 0. For some time I had no idea how was it possible :)
\n\nToday I realized that the problem was with feature generation. I have made a couple of features showing statistics for types of the whole train/test set. These features are okay, but hardly useful!
\n\nAnd now I try a better approach: creating features based on molecules.
\n\nI have little knowledge on physics and chemistry (didn't use them since school), so I use a brute force approach and create features like molecule_type_dist_min - minimum distance per type in molecules.
Here is the link to my kernel which gives -0.210 on public leaderboard.\nThe code for generating all features is provided but isn't used because it creates 450+ features, and thus will exceed kernel memory limit. Instead I have generated all the features locally, run a fast feature selection and created ~50 top features in my kernel.
Of course dumbly generating hundreds of features isn't smart. But I hope that my approach will inspire other people to do better things :)
", "messages": [ { "id": 543213, "postDate": "2019-06-04T11:33:02.047Z", "content": "I have published my first kernel in this competition on the day it started and it had quite a good score. But quite soon we could see scores lower than 0. For some time I had no idea how was it possible :)
\n\nToday I realized that the problem was with feature generation. I have made a couple of features showing statistics for types of the whole train/test set. These features are okay, but hardly useful!
\n\nAnd now I try a better approach: creating features based on molecules.
\n\nI have little knowledge on physics and chemistry (didn't use them since school), so I use a brute force approach and create features like molecule_type_dist_min - minimum distance per type in molecules.
Here is the link to my kernel which gives -0.210 on public leaderboard.\nThe code for generating all features is provided but isn't used because it creates 450+ features, and thus will exceed kernel memory limit. Instead I have generated all the features locally, run a fast feature selection and created ~50 top features in my kernel.
Of course dumbly generating hundreds of features isn't smart. But I hope that my approach will inspire other people to do better things :)
", "rawMarkdown": "I have published my first kernel in this competition on the day it started and it had quite a good score. But quite soon we could see scores lower than 0. For some time I had no idea how was it possible :)\n\nToday I realized that the problem was with feature generation. I have made a couple of features showing statistics for types of the whole train/test set. These features are okay, but hardly useful!\n\nAnd now I try a better approach: creating features based on molecules.\n\nI have little knowledge on physics and chemistry (didn't use them since school), so I use a brute force approach and create features like `molecule_type_dist_min` - minimum distance per type in molecules.\n\nHere is the link to my [kernel](https://www.kaggle.com/artgor/brute-force-feature-engineering) which gives `-0.210` on public leaderboard.\nThe code for generating all features is provided but isn't used because it creates 450+ features, and thus will exceed kernel memory limit. Instead I have generated all the features locally, run a fast feature selection and created ~50 top features in my kernel.\n\nOf course dumbly generating hundreds of features isn't smart. But I hope that my approach will inspire other people to do better things :)", "votes": 59 }, { "id": 570547, "postDate": "2019-07-08T12:44:05.170Z", "content": "great!
", "rawMarkdown": "great!", "votes": 3 }, { "id": 584003, "postDate": "2019-07-25T09:46:55.740Z", "content": "Great
", "rawMarkdown": "Great", "votes": 1 }, { "id": 573082, "postDate": "2019-07-11T18:56:18.800Z", "content": "Great Kernel!
", "rawMarkdown": "Great Kernel!", "votes": 1 }, { "id": 566619, "postDate": "2019-07-02T11:30:05.950Z", "content": "Hi, thanks for sharing .Which technique you used for fast feature selection?
", "rawMarkdown": "Hi, thanks for sharing .Which technique you used for fast feature selection?", "votes": 1, "replies": [ { "id": 566633, "postDate": "2019-07-02T11:46:45.610Z", "content": "I used lgbm feature importance and took top-50 features.
", "rawMarkdown": "I used lgbm feature importance and took top-50 features." } ] }, { "id": 544375, "postDate": "2019-06-05T12:56:09.340Z", "content": "Awesome stuff! Very insightful.
", "rawMarkdown": "Awesome stuff! Very insightful. ", "votes": 1 }, { "id": 543412, "postDate": "2019-06-04T14:15:08.770Z", "content": "Very good Kernel. I'm learning a lot with it. Thanks for sharing
", "rawMarkdown": "Very good Kernel. I'm learning a lot with it. Thanks for sharing", "votes": 1 }, { "id": 544189, "postDate": "2019-06-05T08:50:29.610Z", "content": "You are the TRUE grand master!
", "rawMarkdown": "You are the TRUE grand master!", "votes": 2 }, { "id": 546839, "postDate": "2019-06-07T00:15:14.857Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 546831, "postDate": "2019-06-07T00:05:01.870Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 565726, "postDate": "2019-07-01T10:08:56.487Z", "content": "Thanks for your great kernel!
", "rawMarkdown": "Thanks for your great kernel!", "votes": 1 }, { "id": 545932, "postDate": "2019-06-06T03:58:21.207Z", "content": "very helpful. Thanks for sharing!
", "rawMarkdown": "very helpful. Thanks for sharing!", "votes": 1 }, { "id": 550676, "postDate": "2019-06-11T22:54:58.100Z", "content": "Thanks for sharing!
", "rawMarkdown": "Thanks for sharing!", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 570547, "author_name": "Krishna Katyal", "author_url": "", "post_date": "2019-07-08T12:44:05.170000", "content": "great!
", "votes": 3, "replies": [] }, { "id": 584003, "author_name": "Data Saiyantist", "author_url": "", "post_date": "2019-07-25T09:46:55.740000", "content": "Great
", "votes": 1, "replies": [] }, { "id": 573082, "author_name": "Adam Lewis", "author_url": "", "post_date": "2019-07-11T18:56:18.800000", "content": "Great Kernel!
", "votes": 1, "replies": [] }, { "id": 566619, "author_name": "Reverse Flash", "author_url": "", "post_date": "2019-07-02T11:30:05.950000", "content": "Hi, thanks for sharing .Which technique you used for fast feature selection?
", "votes": 1, "replies": [ { "id": 566633, "author_name": "Andrey Lukyanenko", "author_url": "", "post_date": "2019-07-02T11:46:45.610000", "content": "I used lgbm feature importance and took top-50 features.
", "votes": 0, "replies": [] } ] }, { "id": 544375, "author_name": "9ofalltrades", "author_url": "", "post_date": "2019-06-05T12:56:09.340000", "content": "Awesome stuff! Very insightful.
", "votes": 1, "replies": [] }, { "id": 543412, "author_name": "Adriano Avelar", "author_url": "", "post_date": "2019-06-04T14:15:08.770000", "content": "Very good Kernel. I'm learning a lot with it. Thanks for sharing
", "votes": 1, "replies": [] }, { "id": 544189, "author_name": "Chanran Kim", "author_url": "", "post_date": "2019-06-05T08:50:29.610000", "content": "You are the TRUE grand master!
", "votes": 2, "replies": [] }, { "id": 546839, "author_name": "", "author_url": "", "post_date": "2019-06-07T00:15:14.857000", "content": "", "votes": 2, "replies": [] }, { "id": 546831, "author_name": "", "author_url": "", "post_date": "2019-06-07T00:05:01.870000", "content": "", "votes": 2, "replies": [] }, { "id": 565726, "author_name": "Emin Ozturk", "author_url": "", "post_date": "2019-07-01T10:08:56.487000", "content": "Thanks for your great kernel!
", "votes": 1, "replies": [] }, { "id": 545932, "author_name": "Xin Zhao", "author_url": "", "post_date": "2019-06-06T03:58:21.207000", "content": "very helpful. Thanks for sharing!
", "votes": 1, "replies": [] }, { "id": 550676, "author_name": "", "author_url": "", "post_date": "2019-06-11T22:54:58.100000", "content": "Thanks for sharing!
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "543213": "I have published my first kernel in this competition on the day it started and it had quite a good score. But quite soon we could see scores lower than 0. For some time I had no idea how was it possible :)\n\nToday I realized that the problem was with feature generation. I have made a couple of features showing statistics for types of the whole train/test set. These features are okay, but hardly useful!\n\nAnd now I try a better approach: creating features based on molecules.\n\nI have little knowledge on physics and chemistry (didn't use them since school), so I use a brute force approach and create features like `molecule_type_dist_min` - minimum distance per type in molecules.\n\nHere is the link to my [kernel](https://www.kaggle.com/artgor/brute-force-feature-engineering) which gives `-0.210` on public leaderboard.\nThe code for generating all features is provided but isn't used because it creates 450+ features, and thus will exceed kernel memory limit. Instead I have generated all the features locally, run a fast feature selection and created ~50 top features in my kernel.\n\nOf course dumbly generating hundreds of features isn't smart. But I hope that my approach will inspire other people to do better things :)", "570547": "great!", "584003": "Great", "573082": "Great Kernel!", "566619": "Hi, thanks for sharing .Which technique you used for fast feature selection?", "544375": "Awesome stuff! Very insightful. ", "543412": "Very good Kernel. I'm learning a lot with it. Thanks for sharing", "544189": "You are the TRUE grand master!", "546839": "", "546831": "", "565726": "Thanks for your great kernel!", "545932": "very helpful. Thanks for sharing!", "550676": "Thanks for sharing!" } }