{
  "id": 584788,
  "title": "Mixed Data or Ensemble model?",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/584788",
  "author_name": "",
  "post_date": "2025-06-16T04:50:48.515559400Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>For this comp, one key caveat is that the test set have half of the data with <code>tof_</code> and <code>thm_</code> features marked NULL.<br>\nSo the key is to identify the best way to make the model more robust during inference. I can think of two ways</p>\n<ol>\n<li>Mixed data: mix complete features with missing features and train a single model.</li>\n<li>Mixed model: one model train with full features, while another model with missing features and ensemble during inference?</li>\n</ol>\n<p>Wonder, which one works best for you? Any other ways to counter the problem?</p>",
  "messages": [
    {
      "id": "3225190",
      "postDate": "06/16/2025 04:50:48",
      "content": "<p>For this comp, one key caveat is that the test set have half of the data with <code>tof_</code> and <code>thm_</code> features marked NULL.<br>\nSo the key is to identify the best way to make the model more robust during inference. I can think of two ways</p>\n<ol>\n<li>Mixed data: mix complete features with missing features and train a single model.</li>\n<li>Mixed model: one model train with full features, while another model with missing features and ensemble during inference?</li>\n</ol>\n<p>Wonder, which one works best for you? Any other ways to counter the problem?</p>",
      "rawMarkdown": "For this comp, one key caveat is that the test set have half of the data with `tof_` and `thm_` features marked NULL.\nSo the key is to identify the best way to make the model more robust during inference. I can think of two ways\n\n1. Mixed data: mix complete features with missing features and train a single model.\n2. Mixed model: one model train with full features, while another model with missing features and ensemble during inference?\n\nWonder, which one works best for you? Any other ways to counter the problem?",
      "votes": null
    },
    {
      "id": "3266068",
      "postDate": "08/08/2025 10:46:46",
      "content": "<p>What worked best for you?</p>",
      "rawMarkdown": "What worked best for you?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3266068,
      "author_name": "bogdanturbal",
      "author_url": "",
      "post_date": "08/08/2025 10:46:46",
      "content": "<p>What worked best for you?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3225190": "For this comp, one key caveat is that the test set have half of the data with `tof_` and `thm_` features marked NULL.\nSo the key is to identify the best way to make the model more robust during inference. I can think of two ways\n\n1. Mixed data: mix complete features with missing features and train a single model.\n2. Mixed model: one model train with full features, while another model with missing features and ensemble during inference?\n\nWonder, which one works best for you? Any other ways to counter the problem?",
    "3266068": "What worked best for you?"
  },
  "source": "meta"
}