{
  "id": 21341,
  "title": "Local or Global problem formulation",
  "url": "/competitions/facebook-v-predicting-check-ins/discussion/21341",
  "author_name": "",
  "post_date": "2016-05-31T14:59:29.010Z",
  "votes": null,
  "comment_count": 2,
  "views": 584,
  "content": "<p>Hi guys,</p>\n\n<p>there is some popularity in using a different model for a coordinate based patch of the data, aka a local formulation. This looks reasonable to me since alot of classes will be only aware at one location (am I right?).</p>\n\n<p>So in general:\nIs the local setup prefered to the global one (one classifer on the whole set)?</p>\n\n<p>Regards,\nSterby</p>",
  "messages": [
    {
      "id": "121993",
      "postDate": "05/31/2016 14:59:29",
      "content": "<p>Hi guys,</p>\n\n<p>there is some popularity in using a different model for a coordinate based patch of the data, aka a local formulation. This looks reasonable to me since alot of classes will be only aware at one location (am I right?).</p>\n\n<p>So in general:\nIs the local setup prefered to the global one (one classifer on the whole set)?</p>\n\n<p>Regards,\nSterby</p>",
      "rawMarkdown": "Hi guys,\r\n\r\nthere is some popularity in using a different model for a coordinate based patch of the data, aka a local formulation. This looks reasonable to me since alot of classes will be only aware at one location (am I right?).\r\n\r\nSo in general:\r\nIs the local setup prefered to the global one (one classifer on the whole set)?\r\n\r\nRegards,\r\nSterby",
      "votes": null
    },
    {
      "id": "121994",
      "postDate": "05/31/2016 15:06:24",
      "content": "<p>Hi, </p>\n\n<p>For me it is but not only for the reason you gave but for making the problem manageable. I don't know a technique able to handle 100k classes at the same time...</p>",
      "rawMarkdown": "Hi, \r\n\r\nFor me it is but not only for the reason you gave but for making the problem manageable. I don't know a technique able to handle 100k classes at the same time...",
      "votes": null
    },
    {
      "id": "122078",
      "postDate": "06/01/2016 05:58:29",
      "content": "<p>If you use a localized approach I would suggest to use <strong>overlapping</strong> rectangles in order to reduce the inflated prediction error near the edges of the coordinate patch. In this case you can train your models using all the observations in the rectangles but only use the model with the nearest rectangle center to make predictions. Depending on your grid size it might also make sense to combine the predictions of multiple &quot;close&quot; models.</p>",
      "rawMarkdown": "If you use a localized approach I would suggest to use **overlapping** rectangles in order to reduce the inflated prediction error near the edges of the coordinate patch. In this case you can train your models using all the observations in the rectangles but only use the model with the nearest rectangle center to make predictions. Depending on your grid size it might also make sense to combine the predictions of multiple \"close\" models.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 121994,
      "author_name": "shdgrao",
      "author_url": "",
      "post_date": "05/31/2016 15:06:24",
      "content": "<p>Hi, </p>\n\n<p>For me it is but not only for the reason you gave but for making the problem manageable. I don't know a technique able to handle 100k classes at the same time...</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 122078,
      "author_name": "tvdwiele",
      "author_url": "",
      "post_date": "06/01/2016 05:58:29",
      "content": "<p>If you use a localized approach I would suggest to use <strong>overlapping</strong> rectangles in order to reduce the inflated prediction error near the edges of the coordinate patch. In this case you can train your models using all the observations in the rectangles but only use the model with the nearest rectangle center to make predictions. Depending on your grid size it might also make sense to combine the predictions of multiple &quot;close&quot; models.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "121993": "Hi guys,\r\n\r\nthere is some popularity in using a different model for a coordinate based patch of the data, aka a local formulation. This looks reasonable to me since alot of classes will be only aware at one location (am I right?).\r\n\r\nSo in general:\r\nIs the local setup prefered to the global one (one classifer on the whole set)?\r\n\r\nRegards,\r\nSterby",
    "121994": "Hi, \r\n\r\nFor me it is but not only for the reason you gave but for making the problem manageable. I don't know a technique able to handle 100k classes at the same time...",
    "122078": "If you use a localized approach I would suggest to use **overlapping** rectangles in order to reduce the inflated prediction error near the edges of the coordinate patch. In this case you can train your models using all the observations in the rectangles but only use the model with the nearest rectangle center to make predictions. Depending on your grid size it might also make sense to combine the predictions of multiple \"close\" models."
  },
  "source": "meta"
}