{
  "id": 197574,
  "title": "Any idea for online training?",
  "url": "/competitions/riiid-test-answer-prediction/discussion/197574",
  "author_name": "",
  "post_date": "2020-11-17T04:20:46.190665800Z",
  "votes": 11,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Curious about online training strategies, I have tested utilizing the <code>refit</code> function in LGB, which do improve the results a little (just a little, guess around 0.0001~0.0002 on LB). Is there any suggestion or idea on that?</p>",
  "messages": [
    {
      "id": "1081416",
      "postDate": "11/17/2020 04:20:46",
      "content": "<p>Curious about online training strategies, I have tested utilizing the <code>refit</code> function in LGB, which do improve the results a little (just a little, guess around 0.0001~0.0002 on LB). Is there any suggestion or idea on that?</p>",
      "rawMarkdown": "Curious about online training strategies, I have tested utilizing the `refit` function in LGB, which do improve the results a little (just a little, guess around 0.0001~0.0002 on LB). Is there any suggestion or idea on that?",
      "votes": null
    },
    {
      "id": "1081825",
      "postDate": "11/17/2020 10:17:26",
      "content": "<p>It's hard to believe that a model can learn something new from 2% of the data. I believe that updating the database with new examples should be enough. But I guess it's dependent on features/input representation you are using…</p>",
      "rawMarkdown": "It's hard to believe that a model can learn something new from 2% of the data. I believe that updating the database with new examples should be enough. But I guess it's dependent on features/input representation you are using...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1081825,
      "author_name": "bacicnikola",
      "author_url": "",
      "post_date": "11/17/2020 10:17:26",
      "content": "<p>It's hard to believe that a model can learn something new from 2% of the data. I believe that updating the database with new examples should be enough. But I guess it's dependent on features/input representation you are using…</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1081416": "Curious about online training strategies, I have tested utilizing the `refit` function in LGB, which do improve the results a little (just a little, guess around 0.0001~0.0002 on LB). Is there any suggestion or idea on that?",
    "1081825": "It's hard to believe that a model can learn something new from 2% of the data. I believe that updating the database with new examples should be enough. But I guess it's dependent on features/input representation you are using..."
  },
  "source": "meta"
}