{
  "id": 350222,
  "title": "Better ensembling methods for correlation loss function",
  "url": "/competitions/open-problems-multimodal/discussion/350222",
  "author_name": "Vladimir Slaykovskiy",
  "post_date": "2022-09-04T19:08:20.097000",
  "votes": 9,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I want to share my thoughts on a better method of ensembling for correlation loss. <br>\nCorrelation loss is insensitive to linear transformations of predictions. The implication is that 2 solutions with similar performance could have very different means and standard deviations which makes standard weighting difficult. </p>\n<p>Here I explain how to avoid this issue in your solutions: <a href=\"https://www.kaggle.com/vslaykovsky/lb-0-855-normalized-ensembles-for-pearson-s-r\" target=\"_blank\">https://www.kaggle.com/vslaykovsky/lb-0-855-normalized-ensembles-for-pearson-s-r</a> </p>",
  "messages": [
    {
      "id": 1926367,
      "postDate": "2022-09-04T19:08:20.097Z",
      "content": "<p>I want to share my thoughts on a better method of ensembling for correlation loss. <br>\nCorrelation loss is insensitive to linear transformations of predictions. The implication is that 2 solutions with similar performance could have very different means and standard deviations which makes standard weighting difficult. </p>\n<p>Here I explain how to avoid this issue in your solutions: <a href=\"https://www.kaggle.com/vslaykovsky/lb-0-855-normalized-ensembles-for-pearson-s-r\" target=\"_blank\">https://www.kaggle.com/vslaykovsky/lb-0-855-normalized-ensembles-for-pearson-s-r</a> </p>",
      "rawMarkdown": "I want to share my thoughts on a better method of ensembling for correlation loss. \nCorrelation loss is insensitive to linear transformations of predictions. The implication is that 2 solutions with similar performance could have very different means and standard deviations which makes standard weighting difficult. \n\nHere I explain how to avoid this issue in your solutions: https://www.kaggle.com/vslaykovsky/lb-0-855-normalized-ensembles-for-pearson-s-r \n",
      "votes": 8
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1926367": "I want to share my thoughts on a better method of ensembling for correlation loss. \nCorrelation loss is insensitive to linear transformations of predictions. The implication is that 2 solutions with similar performance could have very different means and standard deviations which makes standard weighting difficult. \n\nHere I explain how to avoid this issue in your solutions: https://www.kaggle.com/vslaykovsky/lb-0-855-normalized-ensembles-for-pearson-s-r \n"
  }
}