{
  "id": 79815,
  "title": "Feature selection improvement",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/79815",
  "author_name": "",
  "post_date": "2019-02-07T19:45:16.172408600Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>How much improvement are you able to get by better feature selection ?</p>",
  "messages": [
    {
      "id": "467825",
      "postDate": "02/07/2019 19:45:16",
      "content": "<p>How much improvement are you able to get by better feature selection ?</p>",
      "rawMarkdown": "How much improvement are you able to get by better feature selection ?",
      "votes": null
    },
    {
      "id": "468179",
      "postDate": "02/08/2019 12:42:53",
      "content": "<p>Hi Manoj.\nThe difference in LB between full set of features and limited set can give you a gain of 0.05 in LB\nFor example:\nFeatures LB\n77 &gt; &gt; 3.723\n195 &gt;&gt; 3.709 best\n267 &gt;&gt; 3.713\n513 &gt;&gt; 3.750 full feature set</p>\n\n<p>I am in the process of narrowing down the best feature list. It is slow on my PC. Two features per day.\n<a href=\"https://www.kaggle.com/kmedvedev/elo-slow-but-bullet-proof-feature-selection\">https://www.kaggle.com/kmedvedev/elo-slow-but-bullet-proof-feature-selection</a></p>",
      "rawMarkdown": "Hi Manoj.\nThe difference in LB between full set of features and limited set can give you a gain of 0.05 in LB\nFor example:\nFeatures LB\n77 &gt; &gt; 3.723\n195 &gt;&gt; 3.709 best\n267 &gt;&gt; 3.713\n513 &gt;&gt; 3.750 full feature set\n\nI am in the process of narrowing down the best feature list. It is slow on my PC. Two features per day.\nhttps://www.kaggle.com/kmedvedev/elo-slow-but-bullet-proof-feature-selection",
      "votes": null
    },
    {
      "id": "469914",
      "postDate": "02/12/2019 03:31:15",
      "content": "<p>I have some success on feature selection with lightgbm feat importance but basically it won't improve a lot in my case (around 0.001 on LB). The main advantage would be to reduce computational burden unless your irrelevant features are really noisy. </p>",
      "rawMarkdown": "I have some success on feature selection with lightgbm feat importance but basically it won't improve a lot in my case (around 0.001 on LB). The main advantage would be to reduce computational burden unless your irrelevant features are really noisy.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 468179,
      "author_name": "kmedvedev",
      "author_url": "",
      "post_date": "02/08/2019 12:42:53",
      "content": "<p>Hi Manoj.\nThe difference in LB between full set of features and limited set can give you a gain of 0.05 in LB\nFor example:\nFeatures LB\n77 &gt; &gt; 3.723\n195 &gt;&gt; 3.709 best\n267 &gt;&gt; 3.713\n513 &gt;&gt; 3.750 full feature set</p>\n\n<p>I am in the process of narrowing down the best feature list. It is slow on my PC. Two features per day.\n<a href=\"https://www.kaggle.com/kmedvedev/elo-slow-but-bullet-proof-feature-selection\">https://www.kaggle.com/kmedvedev/elo-slow-but-bullet-proof-feature-selection</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 469914,
      "author_name": "pukkinming",
      "author_url": "",
      "post_date": "02/12/2019 03:31:15",
      "content": "<p>I have some success on feature selection with lightgbm feat importance but basically it won't improve a lot in my case (around 0.001 on LB). The main advantage would be to reduce computational burden unless your irrelevant features are really noisy. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "467825": "How much improvement are you able to get by better feature selection ?",
    "468179": "Hi Manoj.\nThe difference in LB between full set of features and limited set can give you a gain of 0.05 in LB\nFor example:\nFeatures LB\n77 &gt; &gt; 3.723\n195 &gt;&gt; 3.709 best\n267 &gt;&gt; 3.713\n513 &gt;&gt; 3.750 full feature set\n\nI am in the process of narrowing down the best feature list. It is slow on my PC. Two features per day.\nhttps://www.kaggle.com/kmedvedev/elo-slow-but-bullet-proof-feature-selection",
    "469914": "I have some success on feature selection with lightgbm feat importance but basically it won't improve a lot in my case (around 0.001 on LB). The main advantage would be to reduce computational burden unless your irrelevant features are really noisy."
  },
  "source": "meta"
}