{
  "id": 183466,
  "title": "Label hierarchy vs log-loss?",
  "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/183466",
  "author_name": "maettes",
  "post_date": "2020-09-16T19:22:44.267000",
  "votes": 0,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I might be confusing some things, but considering that we are to predict probabilities for each class, how are we supposed to additionally ensure that the (non-probabilistic) label hierarchy is enforced? The rules state that \"your predictions must adhere to the expected label hierarchy\", but if we predict e.g. <code>0.7</code> for <code>negative_exam_for_pe</code>, how does that translate downstream in the label hierarchy exactly? Does it just mean that e.g. <code>negative_exam_for_pe</code> + <code>indeterminate</code> + [at least one PE] = 1? But then [at least one PE] is again complicated. Would be nice if someone could clear that up :)</p>",
  "messages": [
    {
      "id": 1013611,
      "postDate": "2020-09-16T20:10:30.927Z",
      "content": "<p>Discussed under \"Welcome to the Competition Thread\" (not sure how to post link).</p>",
      "rawMarkdown": "Discussed under \"Welcome to the Competition Thread\" (not sure how to post link).",
      "votes": 1,
      "replies": [
        {
          "id": 1013647,
          "postDate": "2020-09-16T21:04:28.173Z",
          "content": "<p>Thanks, completely missed that!</p>",
          "rawMarkdown": "Thanks, completely missed that!"
        }
      ]
    },
    {
      "id": 1013574,
      "postDate": "2020-09-16T19:22:44.267Z",
      "content": "<p>I might be confusing some things, but considering that we are to predict probabilities for each class, how are we supposed to additionally ensure that the (non-probabilistic) label hierarchy is enforced? The rules state that \"your predictions must adhere to the expected label hierarchy\", but if we predict e.g. <code>0.7</code> for <code>negative_exam_for_pe</code>, how does that translate downstream in the label hierarchy exactly? Does it just mean that e.g. <code>negative_exam_for_pe</code> + <code>indeterminate</code> + [at least one PE] = 1? But then [at least one PE] is again complicated. Would be nice if someone could clear that up :)</p>",
      "rawMarkdown": "I might be confusing some things, but considering that we are to predict probabilities for each class, how are we supposed to additionally ensure that the (non-probabilistic) label hierarchy is enforced? The rules state that \"your predictions must adhere to the expected label hierarchy\", but if we predict e.g. `0.7` for `negative_exam_for_pe`, how does that translate downstream in the label hierarchy exactly? Does it just mean that e.g. `negative_exam_for_pe` + `indeterminate` + [at least one PE] = 1? But then [at least one PE] is again complicated. Would be nice if someone could clear that up :)"
    }
  ],
  "comments": [
    {
      "id": 1013611,
      "author_name": "quadcore/Richard Epstein",
      "author_url": "",
      "post_date": "2020-09-16T20:10:30.927000",
      "content": "<p>Discussed under \"Welcome to the Competition Thread\" (not sure how to post link).</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1013647,
          "author_name": "maettes",
          "author_url": "",
          "post_date": "2020-09-16T21:04:28.173000",
          "content": "<p>Thanks, completely missed that!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1013611": "Discussed under \"Welcome to the Competition Thread\" (not sure how to post link).",
    "1013574": "I might be confusing some things, but considering that we are to predict probabilities for each class, how are we supposed to additionally ensure that the (non-probabilistic) label hierarchy is enforced? The rules state that \"your predictions must adhere to the expected label hierarchy\", but if we predict e.g. `0.7` for `negative_exam_for_pe`, how does that translate downstream in the label hierarchy exactly? Does it just mean that e.g. `negative_exam_for_pe` + `indeterminate` + [at least one PE] = 1? But then [at least one PE] is again complicated. Would be nice if someone could clear that up :)"
  }
}