{
  "id": 20755,
  "title": "What is the best way to ensemble different results?",
  "url": "/competitions/state-farm-distracted-driver-detection/discussion/20755",
  "author_name": "june",
  "post_date": "2016-05-06T04:04:46.217000",
  "votes": 1,
  "comment_count": 0,
  "views": 367,
  "content": "<p>I have gone through the famous <a href=\"http://mlwave.com/kaggle-ensembling-guide/\">kaggle ensemble guide</a>, but I am still not sure how to ensemble for multi class log-loss. The current ensemble idea is from @ZFTurbo - use cross validation and take average. </p>\n\n<p>What I tried was like the following:</p>\n\n<ol>\n<li>vote for the class;</li>\n<li>if the mode exists, select the one prediction with highest probability;</li>\n<li>otherwise, take average</li>\n</ol>\n\n<p>Edited: \nAnother idea is to use weighted average based on validation error. The higher validation error is the lower the weight will be.</p>\n\n<p>But it seemed it didn't work well (at least on the LB). Any ideas or suggestions? </p>",
  "messages": [
    {
      "id": 118919,
      "postDate": "2016-05-06T04:04:46.217Z",
      "content": "<p>I have gone through the famous <a href=\"http://mlwave.com/kaggle-ensembling-guide/\">kaggle ensemble guide</a>, but I am still not sure how to ensemble for multi class log-loss. The current ensemble idea is from @ZFTurbo - use cross validation and take average. </p>\n\n<p>What I tried was like the following:</p>\n\n<ol>\n<li>vote for the class;</li>\n<li>if the mode exists, select the one prediction with highest probability;</li>\n<li>otherwise, take average</li>\n</ol>\n\n<p>Edited: \nAnother idea is to use weighted average based on validation error. The higher validation error is the lower the weight will be.</p>\n\n<p>But it seemed it didn't work well (at least on the LB). Any ideas or suggestions? </p>",
      "rawMarkdown": "I have gone through the famous [kaggle ensemble guide][1], but I am still not sure how to ensemble for multi class log-loss. The current ensemble idea is from @ZFTurbo - use cross validation and take average. \r\n\r\nWhat I tried was like the following:\r\n\r\n1. vote for the class;\r\n2. if the mode exists, select the one prediction with highest probability;\r\n3. otherwise, take average\r\n\r\nEdited: \r\nAnother idea is to use weighted average based on validation error. The higher validation error is the lower the weight will be.\r\n\r\nBut it seemed it didn't work well (at least on the LB). Any ideas or suggestions? \r\n\r\n  [1]: http://mlwave.com/kaggle-ensembling-guide/"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "118919": "I have gone through the famous [kaggle ensemble guide][1], but I am still not sure how to ensemble for multi class log-loss. The current ensemble idea is from @ZFTurbo - use cross validation and take average. \r\n\r\nWhat I tried was like the following:\r\n\r\n1. vote for the class;\r\n2. if the mode exists, select the one prediction with highest probability;\r\n3. otherwise, take average\r\n\r\nEdited: \r\nAnother idea is to use weighted average based on validation error. The higher validation error is the lower the weight will be.\r\n\r\nBut it seemed it didn't work well (at least on the LB). Any ideas or suggestions? \r\n\r\n  [1]: http://mlwave.com/kaggle-ensembling-guide/"
  }
}