{
  "id": 582633,
  "title": "Using Non-BFRB data",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/582633",
  "author_name": "NMVR",
  "post_date": "2025-06-01T15:28:05.925000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>So, the test set contains only the 8 BFRB gestures, however we still have 10 non-BFRB classes, which comprise about 41% of the total sample size.<br>\nHas anyone figured out how they can be used?<br>\nLooking at the leaderboard, the we had a few 70+ on day 1,  while the majority was at a local optima of +-52<br>\nThere seems to be more to it than juts tunning catboost models or trying 1D CNNS / Rocket.</p>\n<p>I would assume these 10 NON-BFRB could be used as proxy labels to some of the \"true\" labels, to increase sparsity and data-size.</p>\n<p>Will try that in the following days, but wondering if anyone has tried that so far?</p>",
  "messages": [
    {
      "id": 3215105,
      "postDate": "2025-06-01T15:28:05.927Z",
      "content": "<p>So, the test set contains only the 8 BFRB gestures, however we still have 10 non-BFRB classes, which comprise about 41% of the total sample size.<br>\nHas anyone figured out how they can be used?<br>\nLooking at the leaderboard, the we had a few 70+ on day 1,  while the majority was at a local optima of +-52<br>\nThere seems to be more to it than juts tunning catboost models or trying 1D CNNS / Rocket.</p>\n<p>I would assume these 10 NON-BFRB could be used as proxy labels to some of the \"true\" labels, to increase sparsity and data-size.</p>\n<p>Will try that in the following days, but wondering if anyone has tried that so far?</p>",
      "rawMarkdown": "So, the test set contains only the 8 BFRB gestures, however we still have 10 non-BFRB classes, which comprise about 41% of the total sample size.\nHas anyone figured out how they can be used?\nLooking at the leaderboard, the we had a few 70+ on day 1,  while the majority was at a local optima of +-52\nThere seems to be more to it than juts tunning catboost models or trying 1D CNNS / Rocket.\n\nI would assume these 10 NON-BFRB could be used as proxy labels to some of the \"true\" labels, to increase sparsity and data-size.\n\nWill try that in the following days, but wondering if anyone has tried that so far?",
      "votes": 1
    },
    {
      "id": 3215222,
      "postDate": "2025-06-01T20:09:30.987Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3215222,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-01T20:09:30.987000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3215105": "So, the test set contains only the 8 BFRB gestures, however we still have 10 non-BFRB classes, which comprise about 41% of the total sample size.\nHas anyone figured out how they can be used?\nLooking at the leaderboard, the we had a few 70+ on day 1,  while the majority was at a local optima of +-52\nThere seems to be more to it than juts tunning catboost models or trying 1D CNNS / Rocket.\n\nI would assume these 10 NON-BFRB could be used as proxy labels to some of the \"true\" labels, to increase sparsity and data-size.\n\nWill try that in the following days, but wondering if anyone has tried that so far?",
    "3215222": ""
  }
}