{
  "id": 12804,
  "title": "Handling the Unknown classes",
  "url": "/competitions/datasciencebowl/discussion/12804",
  "author_name": "",
  "post_date": "2015-03-13T20:36:50.827Z",
  "votes": 1,
  "comment_count": 2,
  "views": 914,
  "content": "<p>Hi all -</p>\n<p>Has anyone played with training their model without the Unknown classes, then post-processing the resulting predictions to shift weight to an appropriate Unknown class if there isn't a clear enough winning predicted class?</p>\n<p>ash</p>",
  "messages": [
    {
      "id": "66147",
      "postDate": "03/13/2015 20:36:50",
      "content": "<p>Hi all -</p>\n<p>Has anyone played with training their model without the Unknown classes, then post-processing the resulting predictions to shift weight to an appropriate Unknown class if there isn't a clear enough winning predicted class?</p>\n<p>ash</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "66161",
      "postDate": "03/14/2015 01:56:45",
      "content": "<p>I have similar idea yesterday , but didn't have time to work it out.&nbsp; My idea is to create one more class Unknown2, eg random noise or other known shape that is not similar to known classes.</p>\n<p>During post processing , Unknown=Unknown+Unknown2</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "66167",
      "postDate": "03/14/2015 08:15:00",
      "content": "<p>It's helps a little. I was able to increase LB by ~0.03 simply by adding 0.001 to all *unknown/*other classes.</p>\n<p>Has anybody tried &quot;smarter&quot; post-processing?</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 66161,
      "author_name": "stevendu",
      "author_url": "",
      "post_date": "03/14/2015 01:56:45",
      "content": "<p>I have similar idea yesterday , but didn't have time to work it out.&nbsp; My idea is to create one more class Unknown2, eg random noise or other known shape that is not similar to known classes.</p>\n<p>During post processing , Unknown=Unknown+Unknown2</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 66167,
      "author_name": "rasimakhunzyanov",
      "author_url": "",
      "post_date": "03/14/2015 08:15:00",
      "content": "<p>It's helps a little. I was able to increase LB by ~0.03 simply by adding 0.001 to all *unknown/*other classes.</p>\n<p>Has anybody tried &quot;smarter&quot; post-processing?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "66147": "",
    "66161": "",
    "66167": ""
  },
  "source": "meta"
}