{
  "id": 11312,
  "title": "Does the sampling strategy matter?",
  "url": "/competitions/avazu-ctr-prediction/discussion/11312",
  "author_name": "",
  "post_date": "2014-12-22T13:56:02.067Z",
  "votes": null,
  "comment_count": 2,
  "views": 1172,
  "content": "<p>In the description, it is said that the click and non-click data is sampled according to <strong>different</strong> strategies. I am wondering whether there exists bias or not.</p>\n<p>More importantly, if the training data and test data are sampled according to different strategies, how can we obtain a reliable prediction?</p>",
  "messages": [
    {
      "id": "61665",
      "postDate": "12/22/2014 13:56:02",
      "content": "<p>In the description, it is said that the click and non-click data is sampled according to <strong>different</strong> strategies. I am wondering whether there exists bias or not.</p>\n<p>More importantly, if the training data and test data are sampled according to different strategies, how can we obtain a reliable prediction?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "61684",
      "postDate": "12/22/2014 18:13:47",
      "content": "<p>I think it simply means the positive and negative instances are sampled with different proportions</p>\n<p>from the real data set, in order to make positive (click) data not so sparse. Typical CTR on real data</p>\n<p>is usually ~1% but we see ~16% from this data.</p>\n\n<p>click vs non-click != training vs test data</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "61728",
      "postDate": "12/23/2014 03:54:09",
      "content": "<p>Thanks for reply. &nbsp;I got it partially.</p>\n<p>I was wondering what will happen when the train data and test data have different pattern?</p>\n<p>Suppose the test data is sampled in a different strategy, say, resulting 10% CTP. Then the knowledge learned from the train data is not applicable for the test data.&nbsp;</p>\n\n<p>[quote=deltap;61684]</p>\n<p>I think it simply means the positive and negative instances are sampled with different proportions</p>\n<p>from the real data set, in order to make positive (click) data not so sparse. Typical CTR on real data</p>\n<p>is usually ~1% but we see ~16% from this data.</p>\n<p>click vs non-click != training vs test data</p>\n<p>[/quote]</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 61684,
      "author_name": "yuxipan",
      "author_url": "",
      "post_date": "12/22/2014 18:13:47",
      "content": "<p>I think it simply means the positive and negative instances are sampled with different proportions</p>\n<p>from the real data set, in order to make positive (click) data not so sparse. Typical CTR on real data</p>\n<p>is usually ~1% but we see ~16% from this data.</p>\n\n<p>click vs non-click != training vs test data</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 61728,
      "author_name": "dataying",
      "author_url": "",
      "post_date": "12/23/2014 03:54:09",
      "content": "<p>Thanks for reply. &nbsp;I got it partially.</p>\n<p>I was wondering what will happen when the train data and test data have different pattern?</p>\n<p>Suppose the test data is sampled in a different strategy, say, resulting 10% CTP. Then the knowledge learned from the train data is not applicable for the test data.&nbsp;</p>\n\n<p>[quote=deltap;61684]</p>\n<p>I think it simply means the positive and negative instances are sampled with different proportions</p>\n<p>from the real data set, in order to make positive (click) data not so sparse. Typical CTR on real data</p>\n<p>is usually ~1% but we see ~16% from this data.</p>\n<p>click vs non-click != training vs test data</p>\n<p>[/quote]</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "61665": "",
    "61684": "",
    "61728": ""
  },
  "source": "meta"
}