{
  "id": 256439,
  "title": "Is There a better Loss Function?",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/256439",
  "author_name": "",
  "post_date": "2021-08-01T16:09:58.763412600Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I've noticed that binary cross-entropy is the norm for most notebooks in the competition since it correlates most with auc score, but this has made me wonder if there is a loss function that is better at maximizing the auc score.</p>",
  "messages": [
    {
      "id": "1407203",
      "postDate": "08/01/2021 16:09:58",
      "content": "<p>I've noticed that binary cross-entropy is the norm for most notebooks in the competition since it correlates most with auc score, but this has made me wonder if there is a loss function that is better at maximizing the auc score.</p>",
      "rawMarkdown": "I've noticed that binary cross-entropy is the norm for most notebooks in the competition since it correlates most with auc score, but this has made me wonder if there is a loss function that is better at maximizing the auc score.",
      "votes": null
    },
    {
      "id": "1449611",
      "postDate": "08/04/2021 23:22:27",
      "content": "<p>Perhaps you can try Focal Loss. TensorFlow has a built-in function for it, while for PyTorch, you will have to create your own or look for implementation made by someone else. </p>\n<p>But I think BCE might still be better for this competition</p>",
      "rawMarkdown": "Perhaps you can try Focal Loss. TensorFlow has a built-in function for it, while for PyTorch, you will have to create your own or look for implementation made by someone else. \n\nBut I think BCE might still be better for this competition",
      "votes": null
    },
    {
      "id": "1468580",
      "postDate": "08/12/2021 13:42:56",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/aristotle609\" target=\"_blank\">@aristotle609</a> ,  </p>\n<p>I think <a href=\"https://towardsdatascience.com/explicit-auc-maximization-70beef6db14e\" target=\"_blank\">this article</a> is useful for you.  <br>\nI'd like to verify this loss function.  </p>\n<p>Thank you!</p>",
      "rawMarkdown": "Hello @aristotle609 ,  \n\nI think [this article](https://towardsdatascience.com/explicit-auc-maximization-70beef6db14e) is useful for you.  \nI'd like to verify this loss function.  \n\nThank you!",
      "votes": null
    },
    {
      "id": "1469642",
      "postDate": "08/13/2021 04:01:10",
      "content": "<p>Thank you <a href=\"https://www.kaggle.com/narikawa\" target=\"_blank\">@narikawa</a> this article is really useful</p>",
      "rawMarkdown": "Thank you @narikawa this article is really useful",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1449611,
      "author_name": "jaechanlee",
      "author_url": "",
      "post_date": "08/04/2021 23:22:27",
      "content": "<p>Perhaps you can try Focal Loss. TensorFlow has a built-in function for it, while for PyTorch, you will have to create your own or look for implementation made by someone else. </p>\n<p>But I think BCE might still be better for this competition</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1468580,
      "author_name": "narikawa",
      "author_url": "",
      "post_date": "08/12/2021 13:42:56",
      "content": "<p>Hello <a href=\"https://www.kaggle.com/aristotle609\" target=\"_blank\">@aristotle609</a> ,  </p>\n<p>I think <a href=\"https://towardsdatascience.com/explicit-auc-maximization-70beef6db14e\" target=\"_blank\">this article</a> is useful for you.  <br>\nI'd like to verify this loss function.  </p>\n<p>Thank you!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1469642,
          "author_name": "aristotle609",
          "author_url": "",
          "post_date": "08/13/2021 04:01:10",
          "content": "<p>Thank you <a href=\"https://www.kaggle.com/narikawa\" target=\"_blank\">@narikawa</a> this article is really useful</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1407203": "I've noticed that binary cross-entropy is the norm for most notebooks in the competition since it correlates most with auc score, but this has made me wonder if there is a loss function that is better at maximizing the auc score.",
    "1449611": "Perhaps you can try Focal Loss. TensorFlow has a built-in function for it, while for PyTorch, you will have to create your own or look for implementation made by someone else. \n\nBut I think BCE might still be better for this competition",
    "1468580": "Hello @aristotle609 ,  \n\nI think [this article](https://towardsdatascience.com/explicit-auc-maximization-70beef6db14e) is useful for you.  \nI'd like to verify this loss function.  \n\nThank you!",
    "1469642": "Thank you @narikawa this article is really useful"
  },
  "source": "meta"
}