{
  "id": 78923,
  "title": "Dose any metrics work well?",
  "url": "/competitions/quora-insincere-questions-classification/discussion/78923",
  "author_name": "",
  "post_date": "2019-01-29T06:46:00.989526600Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I used auc or f1 or val loss as metric , but when I got a better score of those metrics and then posted the code , the LB score was worse . I was so confuse that I don't know how to adjust the model . Dose anyone has a good idea which the metric can guide the right direction of model training ？</p>",
  "messages": [
    {
      "id": "462958",
      "postDate": "01/29/2019 06:46:00",
      "content": "<p>I used auc or f1 or val loss as metric , but when I got a better score of those metrics and then posted the code , the LB score was worse . I was so confuse that I don't know how to adjust the model . Dose anyone has a good idea which the metric can guide the right direction of model training ？</p>",
      "rawMarkdown": "I used auc or f1 or val loss as metric , but when I got a better score of those metrics and then posted the code , the LB score was worse . I was so confuse that I don't know how to adjust the model . Dose anyone has a good idea which the metric can guide the right direction of model training ？",
      "votes": null
    },
    {
      "id": "463324",
      "postDate": "01/29/2019 20:01:10",
      "content": "<p>All the public kernels seem to overfit the data onto the training dataset. Looking at positive accuracy (recall) and negative accuracy on validation set seems to give me better results. Improving negative accuracy keeping recall constant is what i am doing as of now. </p>",
      "rawMarkdown": "All the public kernels seem to overfit the data onto the training dataset. Looking at positive accuracy (recall) and negative accuracy on validation set seems to give me better results. Improving negative accuracy keeping recall constant is what i am doing as of now.",
      "votes": null
    },
    {
      "id": "463700",
      "postDate": "01/30/2019 13:00:42",
      "content": "<p>Thank you for your reply ~ let me try you idea</p>",
      "rawMarkdown": "Thank you for your reply ~ let me try you idea",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 463324,
      "author_name": "karthikkolli0",
      "author_url": "",
      "post_date": "01/29/2019 20:01:10",
      "content": "<p>All the public kernels seem to overfit the data onto the training dataset. Looking at positive accuracy (recall) and negative accuracy on validation set seems to give me better results. Improving negative accuracy keeping recall constant is what i am doing as of now. </p>",
      "votes": null,
      "replies": [
        {
          "id": 463700,
          "author_name": "cloudly1234",
          "author_url": "",
          "post_date": "01/30/2019 13:00:42",
          "content": "<p>Thank you for your reply ~ let me try you idea</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "462958": "I used auc or f1 or val loss as metric , but when I got a better score of those metrics and then posted the code , the LB score was worse . I was so confuse that I don't know how to adjust the model . Dose anyone has a good idea which the metric can guide the right direction of model training ？",
    "463324": "All the public kernels seem to overfit the data onto the training dataset. Looking at positive accuracy (recall) and negative accuracy on validation set seems to give me better results. Improving negative accuracy keeping recall constant is what i am doing as of now.",
    "463700": "Thank you for your reply ~ let me try you idea"
  },
  "source": "meta"
}