{
  "id": 186211,
  "title": "Using competition metric as the loss function",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/186211",
  "author_name": "",
  "post_date": "2020-09-23T16:35:25.521180700Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I have seen many notebooks combine their pinball loss and laplace log likelihood metric as a loss function.<br>\nBut when i tried using just the laplace log likelihood as a loss function without combining it with any pinball loss, my model accuracy remained constant and it did not learn anything<br>\nDoes anyone have an idea as to why that happened</p>",
  "messages": [
    {
      "id": "1024096",
      "postDate": "09/23/2020 16:35:25",
      "content": "<p>I have seen many notebooks combine their pinball loss and laplace log likelihood metric as a loss function.<br>\nBut when i tried using just the laplace log likelihood as a loss function without combining it with any pinball loss, my model accuracy remained constant and it did not learn anything<br>\nDoes anyone have an idea as to why that happened</p>",
      "rawMarkdown": "I have seen many notebooks combine their pinball loss and laplace log likelihood metric as a loss function.\nBut when i tried using just the laplace log likelihood as a loss function without combining it with any pinball loss, my model accuracy remained constant and it did not learn anything\nDoes anyone have an idea as to why that happened",
      "votes": null
    },
    {
      "id": "1026989",
      "postDate": "09/25/2020 18:05:04",
      "content": "<p>There's this previous discussion that addresses it: <a href=\"https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/181505\" target=\"_blank\">https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/181505</a></p>\n<p>In essence, the absolute error term of the loss is hard to deal with, but in principle it should be possible to implement something, but it with many popular frameworks it may require a low-level intervention in the algorithm.</p>",
      "rawMarkdown": "There's this previous discussion that addresses it: https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/181505\n\nIn essence, the absolute error term of the loss is hard to deal with, but in principle it should be possible to implement something, but it with many popular frameworks it may require a low-level intervention in the algorithm.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1026989,
      "author_name": "bjoernholzhauer",
      "author_url": "",
      "post_date": "09/25/2020 18:05:04",
      "content": "<p>There's this previous discussion that addresses it: <a href=\"https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/181505\" target=\"_blank\">https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/181505</a></p>\n<p>In essence, the absolute error term of the loss is hard to deal with, but in principle it should be possible to implement something, but it with many popular frameworks it may require a low-level intervention in the algorithm.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1024096": "I have seen many notebooks combine their pinball loss and laplace log likelihood metric as a loss function.\nBut when i tried using just the laplace log likelihood as a loss function without combining it with any pinball loss, my model accuracy remained constant and it did not learn anything\nDoes anyone have an idea as to why that happened",
    "1026989": "There's this previous discussion that addresses it: https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/181505\n\nIn essence, the absolute error term of the loss is hard to deal with, but in principle it should be possible to implement something, but it with many popular frameworks it may require a low-level intervention in the algorithm."
  },
  "source": "meta"
}