{
  "id": 160671,
  "title": "11th hour hail mary : A new AUC loss function you're free to try.",
  "url": "/competitions/jigsaw-multilingual-toxic-comment-classification/discussion/160671",
  "author_name": "",
  "post_date": "2020-06-22T08:10:18.041913900Z",
  "votes": 10,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hey Kagglers!</p>\n\n<p>I'm not in this competition but I did some recent work on a loss function which\nmore directly targets AUC/ROC than BxE does.   In early tests, for most model\ntypes it outperforms BxE, reaching a higher top score, overfitting less, and doesn't have BxE's habit of overfitting to the point that training score diverges from validation, getting worse with more epochs.</p>\n\n<p>I know it's sort of late in the competition, but maybe a straggler wants to give it a whirl, or even a team that's just a few places from gold and feeling adventurous.  I regret I didn't post this sooner.</p>\n\n<p>Kaggle's weird about running jupyter notebooks right, so here is just a naked script that tests the loss function.  </p>\n\n<p><a href=\"https://www.kaggle.com/iridiumblue/roc-star-an-auc-loss-function-to-challenge-bxe\">Kernel script</a></p>\n\n<p>Here's a paper on the function itself, <a href=\"https://github.com/iridiumblue/roc-star\">Roc-star : An objective function for ROC-AUC that actually works</a> .</p>\n\n<p>It's a great race, Kagglers.   Pace yourselves, take breaks, and win or lose - take joy in the running.</p>",
  "messages": [
    {
      "id": "896492",
      "postDate": "06/22/2020 08:10:18",
      "content": "<p>Hey Kagglers!</p>\n\n<p>I'm not in this competition but I did some recent work on a loss function which\nmore directly targets AUC/ROC than BxE does.   In early tests, for most model\ntypes it outperforms BxE, reaching a higher top score, overfitting less, and doesn't have BxE's habit of overfitting to the point that training score diverges from validation, getting worse with more epochs.</p>\n\n<p>I know it's sort of late in the competition, but maybe a straggler wants to give it a whirl, or even a team that's just a few places from gold and feeling adventurous.  I regret I didn't post this sooner.</p>\n\n<p>Kaggle's weird about running jupyter notebooks right, so here is just a naked script that tests the loss function.  </p>\n\n<p><a href=\"https://www.kaggle.com/iridiumblue/roc-star-an-auc-loss-function-to-challenge-bxe\">Kernel script</a></p>\n\n<p>Here's a paper on the function itself, <a href=\"https://github.com/iridiumblue/roc-star\">Roc-star : An objective function for ROC-AUC that actually works</a> .</p>\n\n<p>It's a great race, Kagglers.   Pace yourselves, take breaks, and win or lose - take joy in the running.</p>",
      "rawMarkdown": "Hey Kagglers!\n\nI'm not in this competition but I did some recent work on a loss function which\nmore directly targets AUC/ROC than BxE does.   In early tests, for most model\ntypes it outperforms BxE, reaching a higher top score, overfitting less, and doesn't have BxE's habit of overfitting to the point that training score diverges from validation, getting worse with more epochs.\n\nI know it's sort of late in the competition, but maybe a straggler wants to give it a whirl, or even a team that's just a few places from gold and feeling adventurous.  I regret I didn't post this sooner.\n\nKaggle's weird about running jupyter notebooks right, so here is just a naked script that tests the loss function.  \n\n[Kernel script](https://www.kaggle.com/iridiumblue/roc-star-an-auc-loss-function-to-challenge-bxe)\n\nHere's a paper on the function itself, [Roc-star : An objective function for ROC-AUC that actually works](https://github.com/iridiumblue/roc-star) .\n\nIt's a great race, Kagglers.   Pace yourselves, take breaks, and win or lose - take joy in the running.",
      "votes": null
    },
    {
      "id": "898551",
      "postDate": "06/23/2020 15:28:30",
      "content": "<p>It seems that the 11th hour hail mary was AUC post processing instead of AUC loss.</p>",
      "rawMarkdown": "It seems that the 11th hour hail mary was AUC post processing instead of AUC loss.",
      "votes": null
    },
    {
      "id": "898971",
      "postDate": "06/23/2020 21:40:46",
      "content": "<p>Hah!  True 'nuf ...</p>",
      "rawMarkdown": "Hah!  True 'nuf ...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 898551,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "06/23/2020 15:28:30",
      "content": "<p>It seems that the 11th hour hail mary was AUC post processing instead of AUC loss.</p>",
      "votes": null,
      "replies": [
        {
          "id": 898971,
          "author_name": "iridiumblue",
          "author_url": "",
          "post_date": "06/23/2020 21:40:46",
          "content": "<p>Hah!  True 'nuf ...</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "896492": "Hey Kagglers!\n\nI'm not in this competition but I did some recent work on a loss function which\nmore directly targets AUC/ROC than BxE does.   In early tests, for most model\ntypes it outperforms BxE, reaching a higher top score, overfitting less, and doesn't have BxE's habit of overfitting to the point that training score diverges from validation, getting worse with more epochs.\n\nI know it's sort of late in the competition, but maybe a straggler wants to give it a whirl, or even a team that's just a few places from gold and feeling adventurous.  I regret I didn't post this sooner.\n\nKaggle's weird about running jupyter notebooks right, so here is just a naked script that tests the loss function.  \n\n[Kernel script](https://www.kaggle.com/iridiumblue/roc-star-an-auc-loss-function-to-challenge-bxe)\n\nHere's a paper on the function itself, [Roc-star : An objective function for ROC-AUC that actually works](https://github.com/iridiumblue/roc-star) .\n\nIt's a great race, Kagglers.   Pace yourselves, take breaks, and win or lose - take joy in the running.",
    "898551": "It seems that the 11th hour hail mary was AUC post processing instead of AUC loss.",
    "898971": "Hah!  True 'nuf ..."
  },
  "source": "meta"
}