{
  "id": 587536,
  "title": "How do you deal with different size in test-features?",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/587536",
  "author_name": "",
  "post_date": "2025-07-01T12:51:34.570989100Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>How do you best deal with the different feature size in the test set? Do you use two seperate models, one for only IMU and one for all? Or do you use one that can handle the missing data?</p>\n<p>This is the first time I'm experiencing such a problem so I'm quite stumped. </p>",
  "messages": [
    {
      "id": "3237993",
      "postDate": "07/01/2025 12:51:34",
      "content": "<p>How do you best deal with the different feature size in the test set? Do you use two seperate models, one for only IMU and one for all? Or do you use one that can handle the missing data?</p>\n<p>This is the first time I'm experiencing such a problem so I'm quite stumped. </p>",
      "rawMarkdown": "How do you best deal with the different feature size in the test set? Do you use two seperate models, one for only IMU and one for all? Or do you use one that can handle the missing data?\n\nThis is the first time I'm experiencing such a problem so I'm quite stumped.",
      "votes": null
    },
    {
      "id": "3238059",
      "postDate": "07/01/2025 13:59:06",
      "content": "<p>Personally, I'd try multiple architectures here - you could build a nice ensemble that's smarter than just averaging predictions.<br>\nWhat I've done before is train both models (IMU-only and full-feature), then stack a meta-learner on top that figures out how to blend them. The cool part is you can feed the IMU model's predictions into the full model as extra features - it's like giving it a \"second opinion\" that often catches patterns you'd miss otherwise.</p>",
      "rawMarkdown": "Personally, I'd try multiple architectures here - you could build a nice ensemble that's smarter than just averaging predictions.\n\nWhat I've done before is train both models (IMU-only and full-feature), then stack a meta-learner on top that figures out how to blend them. The cool part is you can feed the IMU model's predictions into the full model as extra features - it's like giving it a \"second opinion\" that often catches patterns you'd miss otherwise.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3238059,
      "author_name": "taylorsamarel",
      "author_url": "",
      "post_date": "07/01/2025 13:59:06",
      "content": "<p>Personally, I'd try multiple architectures here - you could build a nice ensemble that's smarter than just averaging predictions.<br>\nWhat I've done before is train both models (IMU-only and full-feature), then stack a meta-learner on top that figures out how to blend them. The cool part is you can feed the IMU model's predictions into the full model as extra features - it's like giving it a \"second opinion\" that often catches patterns you'd miss otherwise.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3237993": "How do you best deal with the different feature size in the test set? Do you use two seperate models, one for only IMU and one for all? Or do you use one that can handle the missing data?\n\nThis is the first time I'm experiencing such a problem so I'm quite stumped.",
    "3238059": "Personally, I'd try multiple architectures here - you could build a nice ensemble that's smarter than just averaging predictions.\n\nWhat I've done before is train both models (IMU-only and full-feature), then stack a meta-learner on top that figures out how to blend them. The cool part is you can feed the IMU model's predictions into the full model as extra features - it's like giving it a \"second opinion\" that often catches patterns you'd miss otherwise."
  },
  "source": "meta"
}