{
  "id": 348995,
  "title": "A question about loss function",
  "url": "/competitions/mayo-clinic-strip-ai/discussion/348995",
  "author_name": "zhehao liang",
  "post_date": "2022-08-30T21:57:20.754000",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi. <br>\nMay I ask a question? </p>\n<p>Why most of the people adopt log_loss/cross entropy to train the model?<br>\nI mean, why not use weighted log loss (which is the same in evaluation in competition)?</p>\n<p>I may get answers by checking the past competitions or designing an experiment, but I have an exam this week, so please excuse me for a simple question like this.</p>",
  "messages": [
    {
      "id": 1920087,
      "postDate": "2022-08-30T21:57:20.753Z",
      "content": "<p>Hi. <br>\nMay I ask a question? </p>\n<p>Why most of the people adopt log_loss/cross entropy to train the model?<br>\nI mean, why not use weighted log loss (which is the same in evaluation in competition)?</p>\n<p>I may get answers by checking the past competitions or designing an experiment, but I have an exam this week, so please excuse me for a simple question like this.</p>",
      "rawMarkdown": "Hi. \nMay I ask a question? \n\nWhy most of the people adopt log_loss/cross entropy to train the model?\nI mean, why not use weighted log loss (which is the same in evaluation in competition)?\n\nI may get answers by checking the past competitions or designing an experiment, but I have an exam this week, so please excuse me for a simple question like this.\n\n",
      "votes": 3
    },
    {
      "id": 1922524,
      "postDate": "2022-09-01T14:49:18.803Z",
      "content": "<p>This might be a good experiment to try! It might work! <br>\nFrom my own experience: This approach (engineering the loss to better fit the competition) can yield a huge boost or on the flipside: A huge dump of the score due to overfitting. </p>\n<p>Worth trying! </p>",
      "rawMarkdown": "This might be a good experiment to try! It might work! \nFrom my own experience: This approach (engineering the loss to better fit the competition) can yield a huge boost or on the flipside: A huge dump of the score due to overfitting. \n\nWorth trying! ",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 1922524,
      "author_name": "The Devastator",
      "author_url": "",
      "post_date": "2022-09-01T14:49:18.803000",
      "content": "<p>This might be a good experiment to try! It might work! <br>\nFrom my own experience: This approach (engineering the loss to better fit the competition) can yield a huge boost or on the flipside: A huge dump of the score due to overfitting. </p>\n<p>Worth trying! </p>",
      "votes": 2,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1920087": "Hi. \nMay I ask a question? \n\nWhy most of the people adopt log_loss/cross entropy to train the model?\nI mean, why not use weighted log loss (which is the same in evaluation in competition)?\n\nI may get answers by checking the past competitions or designing an experiment, but I have an exam this week, so please excuse me for a simple question like this.\n\n",
    "1922524": "This might be a good experiment to try! It might work! \nFrom my own experience: This approach (engineering the loss to better fit the competition) can yield a huge boost or on the flipside: A huge dump of the score due to overfitting. \n\nWorth trying! "
  }
}