{
  "id": 5689,
  "title": "Comparing curves. Is it appropriate model? [sanity check]",
  "url": "/competitions/accelerometer-biometric-competition/discussion/5689",
  "author_name": "",
  "post_date": "2013-09-07T15:53:15.800Z",
  "votes": null,
  "comment_count": 5,
  "views": 1966,
  "content": "<p>How do you guys find the right model for this contest? It's very different to the ones I already tried on kaggle and it made me stuck a little. ( I tried only 101 contests and amazon's so far).</p>\n<p>The first model that comes to my mind is comparing the curves. The set of movements is a 3D curve that can be compared to another curve to find the similarity score. The problem here, that in the test sequence there are 300 points, whereas in the training set there are about 500k points for a device. </p>\n<p>I thought to iterate the smaller curve along the bigger curve and see whether there's a point from where curves are similar enough, but it looks like it's gonna be very slow.</p>\n<p>Is it a completely wrong way I'm following?</p>",
  "messages": [
    {
      "id": "30430",
      "postDate": "09/07/2013 15:53:15",
      "content": "<p>How do you guys find the right model for this contest? It's very different to the ones I already tried on kaggle and it made me stuck a little. ( I tried only 101 contests and amazon's so far).</p>\n<p>The first model that comes to my mind is comparing the curves. The set of movements is a 3D curve that can be compared to another curve to find the similarity score. The problem here, that in the test sequence there are 300 points, whereas in the training set there are about 500k points for a device. </p>\n<p>I thought to iterate the smaller curve along the bigger curve and see whether there's a point from where curves are similar enough, but it looks like it's gonna be very slow.</p>\n<p>Is it a completely wrong way I'm following?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "30431",
      "postDate": "09/07/2013 16:06:00",
      "content": "<p>how about calculating the probability that the test sequence was generated by the distribution &quot;represented&quot; by the training sequence? It's not that accurate, and I haven't tried it, but if the training sequence is long enough, its histogram, or rather its smoothed histogram may represent well enough the distribution that generated it.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "30432",
      "postDate": "09/07/2013 16:19:17",
      "content": "<p>Thanks, I'll investigate this idea. How would you make a histogram for this data set? Some N-dimensional histogram or maybe take mean of X, Y, Z?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "30433",
      "postDate": "09/07/2013 16:24:17",
      "content": "<p>I always had problems working in multidimensional spaces, I'll do a different histogram for each axis :)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "30623",
      "postDate": "09/10/2013 00:22:49",
      "content": "<p>I don't know that I would recommend viewing the data in this way.... &nbsp;Histograms of 1-variable are already 2-dimensional.. &nbsp;2-variable histograms are 3-dimensional and extremely difficult to interpret. &nbsp;I used a 3-dimensional plot to investigate the data, but found that it was hard to analyze it in any useful way.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "30624",
      "postDate": "09/10/2013 00:31:05",
      "content": "<p>oops...</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 30431,
      "author_name": "roulinir",
      "author_url": "",
      "post_date": "09/07/2013 16:06:00",
      "content": "<p>how about calculating the probability that the test sequence was generated by the distribution &quot;represented&quot; by the training sequence? It's not that accurate, and I haven't tried it, but if the training sequence is long enough, its histogram, or rather its smoothed histogram may represent well enough the distribution that generated it.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 30432,
      "author_name": "yankov",
      "author_url": "",
      "post_date": "09/07/2013 16:19:17",
      "content": "<p>Thanks, I'll investigate this idea. How would you make a histogram for this data set? Some N-dimensional histogram or maybe take mean of X, Y, Z?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 30433,
      "author_name": "roulinir",
      "author_url": "",
      "post_date": "09/07/2013 16:24:17",
      "content": "<p>I always had problems working in multidimensional spaces, I'll do a different histogram for each axis :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 30623,
      "author_name": "ryanjones",
      "author_url": "",
      "post_date": "09/10/2013 00:22:49",
      "content": "<p>I don't know that I would recommend viewing the data in this way.... &nbsp;Histograms of 1-variable are already 2-dimensional.. &nbsp;2-variable histograms are 3-dimensional and extremely difficult to interpret. &nbsp;I used a 3-dimensional plot to investigate the data, but found that it was hard to analyze it in any useful way.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 30624,
      "author_name": "ryanjones",
      "author_url": "",
      "post_date": "09/10/2013 00:31:05",
      "content": "<p>oops...</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "30430": "",
    "30431": "",
    "30432": "",
    "30433": "",
    "30623": "",
    "30624": ""
  },
  "source": "meta"
}