{
  "id": 655298,
  "title": "imbalance issue",
  "url": "/competitions/recodai-luc-scientific-image-forgery-detection/discussion/655298",
  "author_name": "",
  "post_date": "2025-12-08T08:35:26.084017Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I've noticed that due to the extreme class imbalance at the pixel level, my model is converging to a local minimum where it predicts \"authentic\" for everything. The positive class (forged pixels) is being completely ignored.</p>\n<p>Does anyone have advice on how to tackle this? </p>",
  "messages": [
    {
      "id": "3367157",
      "postDate": "12/08/2025 08:35:26",
      "content": "<p>I've noticed that due to the extreme class imbalance at the pixel level, my model is converging to a local minimum where it predicts \"authentic\" for everything. The positive class (forged pixels) is being completely ignored.</p>\n<p>Does anyone have advice on how to tackle this? </p>",
      "rawMarkdown": "I've noticed that due to the extreme class imbalance at the pixel level, my model is converging to a local minimum where it predicts \"authentic\" for everything. The positive class (forged pixels) is being completely ignored.\n\nDoes anyone have advice on how to tackle this?",
      "votes": null
    },
    {
      "id": "3367630",
      "postDate": "12/08/2025 15:43:48",
      "content": "<p>try to use loss functions that are less sensitive to the imbalance problem (dice loss, focal loss, etc)</p>",
      "rawMarkdown": "try to use loss functions that are less sensitive to the imbalance problem (dice loss, focal loss, etc)",
      "votes": null
    },
    {
      "id": "3380918",
      "postDate": "12/23/2025 11:07:11",
      "content": "<p>Use focal or dice loss</p>",
      "rawMarkdown": "Use focal or dice loss",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3367630,
      "author_name": "mohamedidrissighalmi",
      "author_url": "",
      "post_date": "12/08/2025 15:43:48",
      "content": "<p>try to use loss functions that are less sensitive to the imbalance problem (dice loss, focal loss, etc)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3380918,
      "author_name": "theodorlu",
      "author_url": "",
      "post_date": "12/23/2025 11:07:11",
      "content": "<p>Use focal or dice loss</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3367157": "I've noticed that due to the extreme class imbalance at the pixel level, my model is converging to a local minimum where it predicts \"authentic\" for everything. The positive class (forged pixels) is being completely ignored.\n\nDoes anyone have advice on how to tackle this?",
    "3367630": "try to use loss functions that are less sensitive to the imbalance problem (dice loss, focal loss, etc)",
    "3380918": "Use focal or dice loss"
  },
  "source": "meta"
}