{
  "id": 70763,
  "title": "Surrogate Functions for Maximizing Precision at the Top",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/70763",
  "author_name": "hengck23",
  "post_date": "2018-11-07T04:43:57.448000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Surrogate Functions for Maximizing Precision at the Top</p>\n\n<p><a href=\"http://proceedings.mlr.press/v37/kar15.pdf\">http://proceedings.mlr.press/v37/kar15.pdf</a></p>\n\n<p>The problem of maximizing precision at the\ntop of a ranked list, often dubbed Precision@k\n(prec@k), finds relevance in myriad learning applications\nsuch as ranking, multi-label classification,\nand learning with severe label imbalance.\nHowever, despite its popularity, there exist significant\ngaps in our understanding of this problem\nand its associated performance measure.\nThe most notable of these is the lack of a convex\nupper bounding surrogate for prec@k. We\nalso lack scalable perceptron and stochastic gradient\ndescent algorithms for optimizing this performance\nmeasure. In this paper we make key\ncontributions in these directions. At the heart\nof our results is a family of truly upper bounding\nsurrogates for prec@k. These surrogates are\nmotivated in a principled manner and enjoy attractive\nproperties such as consistency to prec@k\nunder various natural margin/noise conditions</p>",
  "messages": [
    {
      "id": 416681,
      "postDate": "2018-11-07T04:43:57.447Z",
      "content": "<p>Surrogate Functions for Maximizing Precision at the Top</p>\n\n<p><a href=\"http://proceedings.mlr.press/v37/kar15.pdf\">http://proceedings.mlr.press/v37/kar15.pdf</a></p>\n\n<p>The problem of maximizing precision at the\ntop of a ranked list, often dubbed Precision@k\n(prec@k), finds relevance in myriad learning applications\nsuch as ranking, multi-label classification,\nand learning with severe label imbalance.\nHowever, despite its popularity, there exist significant\ngaps in our understanding of this problem\nand its associated performance measure.\nThe most notable of these is the lack of a convex\nupper bounding surrogate for prec@k. We\nalso lack scalable perceptron and stochastic gradient\ndescent algorithms for optimizing this performance\nmeasure. In this paper we make key\ncontributions in these directions. At the heart\nof our results is a family of truly upper bounding\nsurrogates for prec@k. These surrogates are\nmotivated in a principled manner and enjoy attractive\nproperties such as consistency to prec@k\nunder various natural margin/noise conditions</p>",
      "rawMarkdown": "Surrogate Functions for Maximizing Precision at the Top\n\nhttp://proceedings.mlr.press/v37/kar15.pdf\n\nThe problem of maximizing precision at the\ntop of a ranked list, often dubbed Precision@k\n(prec@k), finds relevance in myriad learning applications\nsuch as ranking, multi-label classification,\nand learning with severe label imbalance.\nHowever, despite its popularity, there exist significant\ngaps in our understanding of this problem\nand its associated performance measure.\nThe most notable of these is the lack of a convex\nupper bounding surrogate for prec@k. We\nalso lack scalable perceptron and stochastic gradient\ndescent algorithms for optimizing this performance\nmeasure. In this paper we make key\ncontributions in these directions. At the heart\nof our results is a family of truly upper bounding\nsurrogates for prec@k. These surrogates are\nmotivated in a principled manner and enjoy attractive\nproperties such as consistency to prec@k\nunder various natural margin/noise conditions",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "416681": "Surrogate Functions for Maximizing Precision at the Top\n\nhttp://proceedings.mlr.press/v37/kar15.pdf\n\nThe problem of maximizing precision at the\ntop of a ranked list, often dubbed Precision@k\n(prec@k), finds relevance in myriad learning applications\nsuch as ranking, multi-label classification,\nand learning with severe label imbalance.\nHowever, despite its popularity, there exist significant\ngaps in our understanding of this problem\nand its associated performance measure.\nThe most notable of these is the lack of a convex\nupper bounding surrogate for prec@k. We\nalso lack scalable perceptron and stochastic gradient\ndescent algorithms for optimizing this performance\nmeasure. In this paper we make key\ncontributions in these directions. At the heart\nof our results is a family of truly upper bounding\nsurrogates for prec@k. These surrogates are\nmotivated in a principled manner and enjoy attractive\nproperties such as consistency to prec@k\nunder various natural margin/noise conditions"
  }
}