{
  "id": 5441,
  "title": "Shift in the AllTaggingInfo timestamps",
  "url": "/competitions/belkin-energy-disaggregation-competition/discussion/5441",
  "author_name": "",
  "post_date": "2013-08-18T22:33:25.003Z",
  "votes": null,
  "comment_count": 2,
  "views": 1348,
  "content": "<p>Dear friends,</p>\n<p>I've been looking at the data for House 1, and I have observed a shift (to the past) in the labelling of most of the events in that house.</p>\n<p>Here is an example:</p>\n<p><img src=\"http://img21.imageshack.us/img21/4853/1l0h.jpg\" alt=\"Spectrogram\" width=\"800\" height=\"601\"></p>\n<p>The green and red lines show the on/off instant according to the AllTaggingInfo file. By a quick visual exploration, it looks like the EM signature takes place approximately 20 secs. later than reported.</p>\n<p>&nbsp;</p>\n<p>This fact cannot be explained by the time uncertainty of the FFT, since it is computed every second, so that the time resolution of the spectrogram is approx. 1 sec.</p>\n<p>I could manually fixe some of the labels, but what worries my now is that this shift also occurs in the test set. Let us imagine that the event in the figure belongs to a test set, and that the blue line shows a time instant contained in SampleSubmission.csv. A good predictor would correctly assign a 0, but the label according to which we are evaluated is a 1!</p>\n<p>I have observed a shift in the range of 5 to 30 secs. in the on and off timestamps of almost all the events in H1.</p>\n<p>Can anyone explain this to me? I hope I made my point clear in spite of my poor English.</p>\n<p>Best,</p>",
  "messages": [
    {
      "id": "28918",
      "postDate": "08/18/2013 22:33:25",
      "content": "<p>Dear friends,</p>\n<p>I've been looking at the data for House 1, and I have observed a shift (to the past) in the labelling of most of the events in that house.</p>\n<p>Here is an example:</p>\n<p><img src=\"http://img21.imageshack.us/img21/4853/1l0h.jpg\" alt=\"Spectrogram\" width=\"800\" height=\"601\"></p>\n<p>The green and red lines show the on/off instant according to the AllTaggingInfo file. By a quick visual exploration, it looks like the EM signature takes place approximately 20 secs. later than reported.</p>\n<p>&nbsp;</p>\n<p>This fact cannot be explained by the time uncertainty of the FFT, since it is computed every second, so that the time resolution of the spectrogram is approx. 1 sec.</p>\n<p>I could manually fixe some of the labels, but what worries my now is that this shift also occurs in the test set. Let us imagine that the event in the figure belongs to a test set, and that the blue line shows a time instant contained in SampleSubmission.csv. A good predictor would correctly assign a 0, but the label according to which we are evaluated is a 1!</p>\n<p>I have observed a shift in the range of 5 to 30 secs. in the on and off timestamps of almost all the events in H1.</p>\n<p>Can anyone explain this to me? I hope I made my point clear in spite of my poor English.</p>\n<p>Best,</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "28936",
      "postDate": "08/19/2013 09:46:51",
      "content": "<p>Hello Jose,</p>\n<p>is it the same issue discussed in this post :&nbsp;<a href=\"https://www.kaggle.com/c/belkin-energy-disaggregation-competition/forums/t/5119/event-off-timestamps-don-t-appear-to-line-up-with-power-changes\">https://www.kaggle.com/c/belkin-energy-disaggregation-competition/forums/t/5119/event-off-timestamps-don-t-appear-to-line-up-with-power-changes</a> ?</p>\n<p>If yes, this thread should give you some answers.</p>\n<p>Regards,</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "29378",
      "postDate": "08/24/2013 15:16:17",
      "content": "<p>Thanks, Christophe, it's been useful. I had missed that thread up.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 28936,
      "author_name": "cbourguignat",
      "author_url": "",
      "post_date": "08/19/2013 09:46:51",
      "content": "<p>Hello Jose,</p>\n<p>is it the same issue discussed in this post :&nbsp;<a href=\"https://www.kaggle.com/c/belkin-energy-disaggregation-competition/forums/t/5119/event-off-timestamps-don-t-appear-to-line-up-with-power-changes\">https://www.kaggle.com/c/belkin-energy-disaggregation-competition/forums/t/5119/event-off-timestamps-don-t-appear-to-line-up-with-power-changes</a> ?</p>\n<p>If yes, this thread should give you some answers.</p>\n<p>Regards,</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 29378,
      "author_name": "joseleiva",
      "author_url": "",
      "post_date": "08/24/2013 15:16:17",
      "content": "<p>Thanks, Christophe, it's been useful. I had missed that thread up.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "28918": "",
    "28936": "",
    "29378": ""
  },
  "source": "meta"
}