{
  "id": 136099,
  "title": "What is the reason of choosing Log loss for scoring?",
  "url": "/competitions/deepfake-detection-challenge/discussion/136099",
  "author_name": "",
  "post_date": "2020-03-17T12:55:53.842453Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I wonder why the organizers decided to use log loss for evaluating our submissions. The class balance is different in train and test set and that affects log loss strongly. Why not AUC or similar metrics, for instance? </p>",
  "messages": [
    {
      "id": "776502",
      "postDate": "03/17/2020 12:55:53",
      "content": "<p>I wonder why the organizers decided to use log loss for evaluating our submissions. The class balance is different in train and test set and that affects log loss strongly. Why not AUC or similar metrics, for instance? </p>",
      "rawMarkdown": "I wonder why the organizers decided to use log loss for evaluating our submissions. The class balance is different in train and test set and that affects log loss strongly. Why not AUC or similar metrics, for instance?",
      "votes": null
    },
    {
      "id": "776550",
      "postDate": "03/17/2020 13:19:43",
      "content": "<p>Yeah I wonder the same. Also we want to detect as many deepfake as possible won't recall be a better metric to judge our model then. I guess the hosts want us to build an overall good model which has the best F1 Score that's why logloss</p>",
      "rawMarkdown": "Yeah I wonder the same. Also we want to detect as many deepfake as possible won't recall be a better metric to judge our model then. I guess the hosts want us to build an overall good model which has the best F1 Score that's why logloss",
      "votes": null
    },
    {
      "id": "785712",
      "postDate": "03/25/2020 09:45:27",
      "content": "<p>In my opinion, accuracy is too rough, which can not help to justify acc 0.8 and 0.8, but logloss does.</p>",
      "rawMarkdown": "In my opinion, accuracy is too rough, which can not help to justify acc 0.8 and 0.8, but logloss does.",
      "votes": null
    },
    {
      "id": "785788",
      "postDate": "03/25/2020 11:50:58",
      "content": "<p>What about ROC AUC?</p>",
      "rawMarkdown": "What about ROC AUC?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 776550,
      "author_name": "tanulsingh077",
      "author_url": "",
      "post_date": "03/17/2020 13:19:43",
      "content": "<p>Yeah I wonder the same. Also we want to detect as many deepfake as possible won't recall be a better metric to judge our model then. I guess the hosts want us to build an overall good model which has the best F1 Score that's why logloss</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 785712,
      "author_name": "fionalxd",
      "author_url": "",
      "post_date": "03/25/2020 09:45:27",
      "content": "<p>In my opinion, accuracy is too rough, which can not help to justify acc 0.8 and 0.8, but logloss does.</p>",
      "votes": null,
      "replies": [
        {
          "id": 785788,
          "author_name": "vpaslay",
          "author_url": "",
          "post_date": "03/25/2020 11:50:58",
          "content": "<p>What about ROC AUC?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "776502": "I wonder why the organizers decided to use log loss for evaluating our submissions. The class balance is different in train and test set and that affects log loss strongly. Why not AUC or similar metrics, for instance?",
    "776550": "Yeah I wonder the same. Also we want to detect as many deepfake as possible won't recall be a better metric to judge our model then. I guess the hosts want us to build an overall good model which has the best F1 Score that's why logloss",
    "785712": "In my opinion, accuracy is too rough, which can not help to justify acc 0.8 and 0.8, but logloss does.",
    "785788": "What about ROC AUC?"
  },
  "source": "meta"
}