{
  "id": 25801,
  "title": "Logloss, auc are not correlated with MAP@?",
  "url": "/competitions/outbrain-click-prediction/discussion/25801",
  "author_name": "",
  "post_date": "2016-11-25T10:31:20.910Z",
  "votes": null,
  "comment_count": 1,
  "views": 314,
  "content": "<p>Because of converting predictions to map format and calculating map score takes some time, I was evaluating performance of my models by logloss and auc. But in my last model evaluation I get my best logloss and auc, however the MAP score decreased down. Did someone also has such an issue also? Is it normal?</p>",
  "messages": [
    {
      "id": "146605",
      "postDate": "11/25/2016 10:31:20",
      "content": "<p>Because of converting predictions to map format and calculating map score takes some time, I was evaluating performance of my models by logloss and auc. But in my last model evaluation I get my best logloss and auc, however the MAP score decreased down. Did someone also has such an issue also? Is it normal?</p>",
      "rawMarkdown": "Because of converting predictions to map format and calculating map score takes some time, I was evaluating performance of my models by logloss and auc. But in my last model evaluation I get my best logloss and auc, however the MAP score decreased down. Did someone also has such an issue also? Is it normal?",
      "votes": null
    },
    {
      "id": "146623",
      "postDate": "11/25/2016 11:48:43",
      "content": "<p>Map cares about ordering of the values, so logloss can be lousy (while ordering is fine), and vice versa\n As for auc, how do you evaluate it? If not per display id, there doesnt need to be a monotone relationship: slightly messed up ordering per display can still give a good auc globally.</p>",
      "rawMarkdown": "Map cares about ordering of the values, so logloss can be lousy (while ordering is fine), and vice versa\r\n As for auc, how do you evaluate it? If not per display id, there doesnt need to be a monotone relationship: slightly messed up ordering per display can still give a good auc globally.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 146623,
      "author_name": "konradb",
      "author_url": "",
      "post_date": "11/25/2016 11:48:43",
      "content": "<p>Map cares about ordering of the values, so logloss can be lousy (while ordering is fine), and vice versa\n As for auc, how do you evaluate it? If not per display id, there doesnt need to be a monotone relationship: slightly messed up ordering per display can still give a good auc globally.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "146605": "Because of converting predictions to map format and calculating map score takes some time, I was evaluating performance of my models by logloss and auc. But in my last model evaluation I get my best logloss and auc, however the MAP score decreased down. Did someone also has such an issue also? Is it normal?",
    "146623": "Map cares about ordering of the values, so logloss can be lousy (while ordering is fine), and vice versa\r\n As for auc, how do you evaluate it? If not per display id, there doesnt need to be a monotone relationship: slightly messed up ordering per display can still give a good auc globally."
  },
  "source": "meta"
}