{
  "id": 16988,
  "title": "what was the trick to beat the ks-test using other files?",
  "url": "/competitions/flavours-of-physics/discussion/16988",
  "author_name": "",
  "post_date": "2015-10-13T22:31:36.143Z",
  "votes": null,
  "comment_count": 3,
  "views": 627,
  "content": "<p>I tried various methods to beat the KS-test using a classifier built on train and check agreement. e.g. 1 if train and 0 if check agreement. I also tried 1 if train and 0 if random permutation of columns in train. Then I tried scaling predictions by the probability of train. I don't think this worked - it didn't allow me to beat the test. </p>\n\n<p>I tried a few different approaches and I was unable to consistently beat the ks-test for a bagged xg model of 30 different XG models. My KS score was still above 0.26+ at times.</p>\n\n<p>I know there are methods of beating the test like 1) blending, 2) removing features like SPD, 3) only using min_ANNmuon&lt;0.4, etc. but was there some way of directly using check agreement to bypass the test entirely?</p>",
  "messages": [
    {
      "id": "96046",
      "postDate": "10/13/2015 22:31:36",
      "content": "<p>I tried various methods to beat the KS-test using a classifier built on train and check agreement. e.g. 1 if train and 0 if check agreement. I also tried 1 if train and 0 if random permutation of columns in train. Then I tried scaling predictions by the probability of train. I don't think this worked - it didn't allow me to beat the test. </p>\n\n<p>I tried a few different approaches and I was unable to consistently beat the ks-test for a bagged xg model of 30 different XG models. My KS score was still above 0.26+ at times.</p>\n\n<p>I know there are methods of beating the test like 1) blending, 2) removing features like SPD, 3) only using min_ANNmuon&lt;0.4, etc. but was there some way of directly using check agreement to bypass the test entirely?</p>",
      "rawMarkdown": "I tried various methods to beat the KS-test using a classifier built on train and check agreement. e.g. 1 if train and 0 if check agreement. I also tried 1 if train and 0 if random permutation of columns in train. Then I tried scaling predictions by the probability of train. I don't think this worked - it didn't allow me to beat the test. \r\n\r\nI tried a few different approaches and I was unable to consistently beat the ks-test for a bagged xg model of 30 different XG models. My KS score was still above 0.26+ at times.\r\n\r\nI know there are methods of beating the test like 1) blending, 2) removing features like SPD, 3) only using min_ANNmuon<0.4, etc. but was there some way of directly using check agreement to bypass the test entirely?",
      "votes": null
    },
    {
      "id": "96059",
      "postDate": "10/14/2015 02:18:57",
      "content": "<p>Fully agree Mr. Phunter below.</p>",
      "rawMarkdown": "Fully agree Mr. Phunter below.",
      "votes": null
    },
    {
      "id": "96144",
      "postDate": "10/14/2015 19:51:03",
      "content": "<p>after the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.</p>",
      "rawMarkdown": "after the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.",
      "votes": null
    },
    {
      "id": "96145",
      "postDate": "10/14/2015 20:03:37",
      "content": "<p>Could you kindly tell me where to find the solutions of the first 28 terms? Thanks. </p>\n\n<p>[quote=phunter;96144]</p>\n\n<p>after the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.</p>\n\n<p>[/quote]</p>",
      "rawMarkdown": "Could you kindly tell me where to find the solutions of the first 28 terms? Thanks. \r\n\r\n[quote=phunter;96144]\r\n\r\nafter the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.\r\n\r\n[/quote]",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 96059,
      "author_name": "",
      "author_url": "",
      "post_date": "10/14/2015 02:18:57",
      "content": "<p>Fully agree Mr. Phunter below.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 96144,
      "author_name": "phunter",
      "author_url": "",
      "post_date": "10/14/2015 19:51:03",
      "content": "<p>after the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 96145,
      "author_name": "liuchengwang",
      "author_url": "",
      "post_date": "10/14/2015 20:03:37",
      "content": "<p>Could you kindly tell me where to find the solutions of the first 28 terms? Thanks. </p>\n\n<p>[quote=phunter;96144]</p>\n\n<p>after the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.</p>\n\n<p>[/quote]</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "96046": "I tried various methods to beat the KS-test using a classifier built on train and check agreement. e.g. 1 if train and 0 if check agreement. I also tried 1 if train and 0 if random permutation of columns in train. Then I tried scaling predictions by the probability of train. I don't think this worked - it didn't allow me to beat the test. \r\n\r\nI tried a few different approaches and I was unable to consistently beat the ks-test for a bagged xg model of 30 different XG models. My KS score was still above 0.26+ at times.\r\n\r\nI know there are methods of beating the test like 1) blending, 2) removing features like SPD, 3) only using min_ANNmuon<0.4, etc. but was there some way of directly using check agreement to bypass the test entirely?",
    "96059": "Fully agree Mr. Phunter below.",
    "96144": "after the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.",
    "96145": "Could you kindly tell me where to find the solutions of the first 28 terms? Thanks. \r\n\r\n[quote=phunter;96144]\r\n\r\nafter the first 28 teams solution, I may give more 'legitimate' tricks, e.g. making xgboost a slight weaker classifier by some parameters and/or feature engineering. P.S. the 3rd trick didn't work, I tried.\r\n\r\n[/quote]"
  },
  "source": "meta"
}