{
  "id": 14633,
  "title": "Did anyone use time series features",
  "url": "/competitions/facebook-recruiting-iv-human-or-bot/discussion/14633",
  "author_name": "",
  "post_date": "2015-06-09T12:16:54.413Z",
  "votes": null,
  "comment_count": 2,
  "views": 929,
  "content": "<p>Hi,</p>\n<p>I did not have time much on this competition, so I mostly focused on all the bid times like linear fit of cumulative bids vs time and some frequency spectrum features with FFT. &nbsp;The most successful features were:</p>\n<ul>\n<li>linear fit slope and y-intercept of cumulative bids over time</li>\n<li>FFT spectrum linear fit y-intercept</li>\n<li>standard deviation of the FFT spectrum</li>\n<li><a href=\"http://en.wikipedia.org/wiki/Spectral_flatness\">spectral flatness</a></li>\n</ul>\n\n<p>Most search I have done, lead to time series extrapolations/predictions, not really classification. &nbsp;Which time series manipulations were most successful for you?</p>\n<p>Cheers!</p>",
  "messages": [
    {
      "id": "81339",
      "postDate": "06/09/2015 12:16:54",
      "content": "<p>Hi,</p>\n<p>I did not have time much on this competition, so I mostly focused on all the bid times like linear fit of cumulative bids vs time and some frequency spectrum features with FFT. &nbsp;The most successful features were:</p>\n<ul>\n<li>linear fit slope and y-intercept of cumulative bids over time</li>\n<li>FFT spectrum linear fit y-intercept</li>\n<li>standard deviation of the FFT spectrum</li>\n<li><a href=\"http://en.wikipedia.org/wiki/Spectral_flatness\">spectral flatness</a></li>\n</ul>\n\n<p>Most search I have done, lead to time series extrapolations/predictions, not really classification. &nbsp;Which time series manipulations were most successful for you?</p>\n<p>Cheers!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81344",
      "postDate": "06/09/2015 12:59:39",
      "content": "<p>Average reaction time to the previous&nbsp;bid and average difference between all bidder's bids grouped by&nbsp;auction. Both features was in top-8 by importance.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "81382",
      "postDate": "06/09/2015 17:52:25",
      "content": "<p>@Artem</p>\n<p>Could you expand a bit more on the time features that you used. I think time was the one field that I had very little success in extracting useful features that might have made the difference between top 25% and top 10% on the LB.&nbsp;</p>\n<p>I tried to engineer an average bid frequency (ie (max time - min time)/ #bids grouped by auction) for each bidder, but that feature seemed to have little or no importance in classification when I used random forests or gradient boosted trees. &nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 81344,
      "author_name": "rushter",
      "author_url": "",
      "post_date": "06/09/2015 12:59:39",
      "content": "<p>Average reaction time to the previous&nbsp;bid and average difference between all bidder's bids grouped by&nbsp;auction. Both features was in top-8 by importance.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 81382,
      "author_name": "vijaysathish",
      "author_url": "",
      "post_date": "06/09/2015 17:52:25",
      "content": "<p>@Artem</p>\n<p>Could you expand a bit more on the time features that you used. I think time was the one field that I had very little success in extracting useful features that might have made the difference between top 25% and top 10% on the LB.&nbsp;</p>\n<p>I tried to engineer an average bid frequency (ie (max time - min time)/ #bids grouped by auction) for each bidder, but that feature seemed to have little or no importance in classification when I used random forests or gradient boosted trees. &nbsp;</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "81339": "",
    "81344": "",
    "81382": ""
  },
  "source": "meta"
}