{
  "id": 15866,
  "title": "How to handle data with no labels at all",
  "url": "/competitions/grasp-and-lift-eeg-detection/discussion/15866",
  "author_name": "",
  "post_date": "2015-08-10T04:42:13.150Z",
  "votes": null,
  "comment_count": 3,
  "views": 433,
  "content": "<p>How do you guys handle with data with no labels for all of the events.</p>\n\n<p>Especially if your are using a neural network with a softmax output, I think it would be hard to model data with no labels.</p>\n\n<p>A possible solution might be using a binary model for all of the six labels. But is this the only way to solve this problem?</p>\n\n<p>How can I deal with this problem?</p>",
  "messages": [
    {
      "id": "89002",
      "postDate": "08/10/2015 04:42:13",
      "content": "<p>How do you guys handle with data with no labels for all of the events.</p>\n\n<p>Especially if your are using a neural network with a softmax output, I think it would be hard to model data with no labels.</p>\n\n<p>A possible solution might be using a binary model for all of the six labels. But is this the only way to solve this problem?</p>\n\n<p>How can I deal with this problem?</p>",
      "rawMarkdown": "How do you guys handle with data with no labels for all of the events.\r\n\r\nEspecially if your are using a neural network with a softmax output, I think it would be hard to model data with no labels.\r\n\r\nA possible solution might be using a binary model for all of the six labels. But is this the only way to solve this problem?\r\n\r\nHow can I deal with this problem?",
      "votes": null
    },
    {
      "id": "89087",
      "postDate": "08/11/2015 08:47:37",
      "content": "<p>I think yes, OneHot coding is a general approach. </p>",
      "rawMarkdown": "I think yes, OneHot coding is a general approach.",
      "votes": null
    },
    {
      "id": "89089",
      "postDate": "08/11/2015 09:33:19",
      "content": "<p>Note that there are also cases with overlapping events, so softmax is not a good model in any case. \nA separate binary output for each of the six labels is most likely what you'd want to do. </p>",
      "rawMarkdown": "Note that there are also cases with overlapping events, so softmax is not a good model in any case. \r\nA separate binary output for each of the six labels is most likely what you'd want to do.",
      "votes": null
    },
    {
      "id": "89294",
      "postDate": "08/13/2015 12:46:26",
      "content": "<p>In case you choose OneHot you have to code the outputs to be OneHot, e.g.</p>\n\n<p>1,0,0,0,0,0 -&gt;1,0,0,0,0,0,0,0\n1,1,0,0,0,0 -&gt; 0,1,0,0,0,0,0,0\n....</p>",
      "rawMarkdown": "In case you choose OneHot you have to code the outputs to be OneHot, e.g.\r\n\r\n1,0,0,0,0,0 ->1,0,0,0,0,0,0,0\r\n1,1,0,0,0,0 -> 0,1,0,0,0,0,0,0\r\n....",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 89087,
      "author_name": "tnilab",
      "author_url": "",
      "post_date": "08/11/2015 08:47:37",
      "content": "<p>I think yes, OneHot coding is a general approach. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 89089,
      "author_name": "qwlouse",
      "author_url": "",
      "post_date": "08/11/2015 09:33:19",
      "content": "<p>Note that there are also cases with overlapping events, so softmax is not a good model in any case. \nA separate binary output for each of the six labels is most likely what you'd want to do. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 89294,
      "author_name": "tnilab",
      "author_url": "",
      "post_date": "08/13/2015 12:46:26",
      "content": "<p>In case you choose OneHot you have to code the outputs to be OneHot, e.g.</p>\n\n<p>1,0,0,0,0,0 -&gt;1,0,0,0,0,0,0,0\n1,1,0,0,0,0 -&gt; 0,1,0,0,0,0,0,0\n....</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "89002": "How do you guys handle with data with no labels for all of the events.\r\n\r\nEspecially if your are using a neural network with a softmax output, I think it would be hard to model data with no labels.\r\n\r\nA possible solution might be using a binary model for all of the six labels. But is this the only way to solve this problem?\r\n\r\nHow can I deal with this problem?",
    "89087": "I think yes, OneHot coding is a general approach.",
    "89089": "Note that there are also cases with overlapping events, so softmax is not a good model in any case. \r\nA separate binary output for each of the six labels is most likely what you'd want to do.",
    "89294": "In case you choose OneHot you have to code the outputs to be OneHot, e.g.\r\n\r\n1,0,0,0,0,0 ->1,0,0,0,0,0,0,0\r\n1,1,0,0,0,0 -> 0,1,0,0,0,0,0,0\r\n...."
  },
  "source": "meta"
}