{
  "id": 100790,
  "title": "Optimized rounder benefit?",
  "url": "/competitions/aptos2019-blindness-detection/discussion/100790",
  "author_name": "",
  "post_date": "2019-07-21T03:18:55.564788400Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Has anyone used an optimized qwk rounder and seen any lb benefits? For me, just using the default coefficients(0.5, 1.5, 2.5, 3.5) work better basically every time.</p>",
  "messages": [
    {
      "id": "580902",
      "postDate": "07/21/2019 03:18:55",
      "content": "<p>Has anyone used an optimized qwk rounder and seen any lb benefits? For me, just using the default coefficients(0.5, 1.5, 2.5, 3.5) work better basically every time.</p>",
      "rawMarkdown": "Has anyone used an optimized qwk rounder and seen any lb benefits? For me, just using the default coefficients(0.5, 1.5, 2.5, 3.5) work better basically every time.",
      "votes": null
    },
    {
      "id": "588110",
      "postDate": "07/30/2019 07:10:11",
      "content": "<p>I tried optimized qwk rounder(borrow from: <a href=\"https://www.kaggle.com/hmendonca/aptos19-regressor-fastai-oversampling-tta\">https://www.kaggle.com/hmendonca/aptos19-regressor-fastai-oversampling-tta</a>),\nit did improve my lb score slightly (from 0.712(using default coef[0.5, 1.5, 2.5, 3.5]) to 0.714)\nIMO, if dataset is balanced(each category has same number of samples), then the default coef should be a ressonable coef;\nelse, adjusting the coef maybe a good choice to improve kappa score.</p>",
      "rawMarkdown": "I tried optimized qwk rounder(borrow from: https://www.kaggle.com/hmendonca/aptos19-regressor-fastai-oversampling-tta),\nit did improve my lb score slightly (from 0.712(using default coef[0.5, 1.5, 2.5, 3.5]) to 0.714)\nIMO, if dataset is balanced(each category has same number of samples), then the default coef should be a ressonable coef;\nelse, adjusting the coef maybe a good choice to improve kappa score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 588110,
      "author_name": "frank518",
      "author_url": "",
      "post_date": "07/30/2019 07:10:11",
      "content": "<p>I tried optimized qwk rounder(borrow from: <a href=\"https://www.kaggle.com/hmendonca/aptos19-regressor-fastai-oversampling-tta\">https://www.kaggle.com/hmendonca/aptos19-regressor-fastai-oversampling-tta</a>),\nit did improve my lb score slightly (from 0.712(using default coef[0.5, 1.5, 2.5, 3.5]) to 0.714)\nIMO, if dataset is balanced(each category has same number of samples), then the default coef should be a ressonable coef;\nelse, adjusting the coef maybe a good choice to improve kappa score.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "580902": "Has anyone used an optimized qwk rounder and seen any lb benefits? For me, just using the default coefficients(0.5, 1.5, 2.5, 3.5) work better basically every time.",
    "588110": "I tried optimized qwk rounder(borrow from: https://www.kaggle.com/hmendonca/aptos19-regressor-fastai-oversampling-tta),\nit did improve my lb score slightly (from 0.712(using default coef[0.5, 1.5, 2.5, 3.5]) to 0.714)\nIMO, if dataset is balanced(each category has same number of samples), then the default coef should be a ressonable coef;\nelse, adjusting the coef maybe a good choice to improve kappa score."
  },
  "source": "meta"
}