{
  "id": 150300,
  "title": "Is there any use of seeing weighted AUC on a batch?",
  "url": "/competitions/alaska2-image-steganalysis/discussion/150300",
  "author_name": "Harsh Jalan",
  "post_date": "2020-05-11T19:44:43.311000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I do not feel that there is any use in including the weighted AUC evaluation metric as a training metric, since the dataset, when considering binary output, is heavily imbalanced and there is a great chance that the true labels in a batch will all be 1, which in turn, makes the false positive rate NaN, since there cannot be a false positive if all true outputs are positive, I believe the weighted AUC only makes sense when we are calculating it on a complete dataset (either test or train). </p>",
  "messages": [
    {
      "id": 843053,
      "postDate": "2020-05-11T19:44:43.310Z",
      "content": "<p>I do not feel that there is any use in including the weighted AUC evaluation metric as a training metric, since the dataset, when considering binary output, is heavily imbalanced and there is a great chance that the true labels in a batch will all be 1, which in turn, makes the false positive rate NaN, since there cannot be a false positive if all true outputs are positive, I believe the weighted AUC only makes sense when we are calculating it on a complete dataset (either test or train). </p>",
      "rawMarkdown": "I do not feel that there is any use in including the weighted AUC evaluation metric as a training metric, since the dataset, when considering binary output, is heavily imbalanced and there is a great chance that the true labels in a batch will all be 1, which in turn, makes the false positive rate NaN, since there cannot be a false positive if all true outputs are positive, I believe the weighted AUC only makes sense when we are calculating it on a complete dataset (either test or train). ",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "843053": "I do not feel that there is any use in including the weighted AUC evaluation metric as a training metric, since the dataset, when considering binary output, is heavily imbalanced and there is a great chance that the true labels in a batch will all be 1, which in turn, makes the false positive rate NaN, since there cannot be a false positive if all true outputs are positive, I believe the weighted AUC only makes sense when we are calculating it on a complete dataset (either test or train). "
  }
}