{
  "id": 9408,
  "title": "Does vw need feature normalization?",
  "url": "/competitions/acquire-valued-shoppers-challenge/discussion/9408",
  "author_name": "",
  "post_date": "2014-06-07T03:06:46.347Z",
  "votes": null,
  "comment_count": 2,
  "views": 1282,
  "content": "<p>From the manual it seems vw can do an active, per feature learning so normalization is unnecessary.</p>\n<p><strong>Am I right ?</strong></p>\n<p>From the examples in its github, I see the features are not normalized neither. &nbsp;</p>",
  "messages": [
    {
      "id": "48816",
      "postDate": "06/07/2014 03:06:46",
      "content": "<p>From the manual it seems vw can do an active, per feature learning so normalization is unnecessary.</p>\n<p><strong>Am I right ?</strong></p>\n<p>From the examples in its github, I see the features are not normalized neither. &nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "48858",
      "postDate": "06/08/2014 10:51:41",
      "content": "<p>VW with&nbsp;basic&nbsp;options are linear, so it's algorithms mostly shouldn't be sensitive to features linear scaling</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "48863",
      "postDate": "06/08/2014 13:33:04",
      "content": "<p>VW standardly normalizes feature values. See the paper&nbsp;http://arxiv.org/pdf/1305.6646.pdf on how they did this with online learning.</p>\n\n<p>You can turn normalization off by specifying the update rules:</p>\n\n<p><code>--normalized --invariant --adaptive</code></p>\n\n<p>Change into:</p>\n\n<p><code>--invariant --adaptive</code></p>\n\n<p>Then you can do your own normalization. My guess it that it won't significantly improve over doing the standard online normalization.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 48858,
      "author_name": "dremovd",
      "author_url": "",
      "post_date": "06/08/2014 10:51:41",
      "content": "<p>VW with&nbsp;basic&nbsp;options are linear, so it's algorithms mostly shouldn't be sensitive to features linear scaling</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 48863,
      "author_name": "triskelion",
      "author_url": "",
      "post_date": "06/08/2014 13:33:04",
      "content": "<p>VW standardly normalizes feature values. See the paper&nbsp;http://arxiv.org/pdf/1305.6646.pdf on how they did this with online learning.</p>\n\n<p>You can turn normalization off by specifying the update rules:</p>\n\n<p><code>--normalized --invariant --adaptive</code></p>\n\n<p>Change into:</p>\n\n<p><code>--invariant --adaptive</code></p>\n\n<p>Then you can do your own normalization. My guess it that it won't significantly improve over doing the standard online normalization.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "48816": "",
    "48858": "",
    "48863": ""
  },
  "source": "meta"
}