{
  "id": 152745,
  "title": "high acc but low f2_score",
  "url": "/competitions/imet-2020-fgvc7/discussion/152745",
  "author_name": "",
  "post_date": "2020-05-21T07:52:30.240858300Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>The model has a high acc (0.998) but a low f2_score(0.006) according training log, I don't know the reason. Can someone gives some advice or helps me to analyse it? Thanks</p>",
  "messages": [
    {
      "id": "855784",
      "postDate": "05/21/2020 07:52:30",
      "content": "<p>The model has a high acc (0.998) but a low f2_score(0.006) according training log, I don't know the reason. Can someone gives some advice or helps me to analyse it? Thanks</p>",
      "rawMarkdown": "The model has a high acc (0.998) but a low f2_score(0.006) according training log, I don't know the reason. Can someone gives some advice or helps me to analyse it? Thanks",
      "votes": null
    },
    {
      "id": "855863",
      "postDate": "05/21/2020 09:32:13",
      "content": "<p>Its because in this task a lot of labels and for every data sample only little part of this labels are positive. There is only 0,127 persent of positive labels in the dataset, so if your model predict only negtive labels for every sample you can get 99,8 perent overall accuracy. I think better to use some other metrics, for example calculate accuracy only for positive labels and amount of false positive labels, it is gives better understanding what model do.</p>",
      "rawMarkdown": "Its because in this task a lot of labels and for every data sample only little part of this labels are positive. There is only 0,127 persent of positive labels in the dataset, so if your model predict only negtive labels for every sample you can get 99,8 perent overall accuracy. I think better to use some other metrics, for example calculate accuracy only for positive labels and amount of false positive labels, it is gives better understanding what model do.",
      "votes": null
    },
    {
      "id": "856044",
      "postDate": "05/21/2020 12:48:11",
      "content": "<p>What's the meaning of positive labels and negtive labels?  Positive labels means labels occur in row of training data? Or positive labels means labels occur in ground truth while predicting testing data? Actually, my submissions include many labels each row.</p>",
      "rawMarkdown": "What's the meaning of positive labels and negtive labels?  Positive labels means labels occur in row of training data? Or positive labels means labels occur in ground truth while predicting testing data? Actually, my submissions include many labels each row.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 855863,
      "author_name": "ragerage",
      "author_url": "",
      "post_date": "05/21/2020 09:32:13",
      "content": "<p>Its because in this task a lot of labels and for every data sample only little part of this labels are positive. There is only 0,127 persent of positive labels in the dataset, so if your model predict only negtive labels for every sample you can get 99,8 perent overall accuracy. I think better to use some other metrics, for example calculate accuracy only for positive labels and amount of false positive labels, it is gives better understanding what model do.</p>",
      "votes": null,
      "replies": [
        {
          "id": 856044,
          "author_name": "datamaters",
          "author_url": "",
          "post_date": "05/21/2020 12:48:11",
          "content": "<p>What's the meaning of positive labels and negtive labels?  Positive labels means labels occur in row of training data? Or positive labels means labels occur in ground truth while predicting testing data? Actually, my submissions include many labels each row.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "855784": "The model has a high acc (0.998) but a low f2_score(0.006) according training log, I don't know the reason. Can someone gives some advice or helps me to analyse it? Thanks",
    "855863": "Its because in this task a lot of labels and for every data sample only little part of this labels are positive. There is only 0,127 persent of positive labels in the dataset, so if your model predict only negtive labels for every sample you can get 99,8 perent overall accuracy. I think better to use some other metrics, for example calculate accuracy only for positive labels and amount of false positive labels, it is gives better understanding what model do.",
    "856044": "What's the meaning of positive labels and negtive labels?  Positive labels means labels occur in row of training data? Or positive labels means labels occur in ground truth while predicting testing data? Actually, my submissions include many labels each row."
  },
  "source": "meta"
}