{
  "id": 114858,
  "title": "it shouldn't be 6 class, but 5 class and hierarchical class",
  "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/114858",
  "author_name": "hengck23",
  "post_date": "2019-10-29T18:00:25.061000",
  "votes": 5,
  "comment_count": 6,
  "views": 0,
  "content": "<p>shouldn't be it be  : \n```\np(any)  = 1 - p(not any) = 1 - (1-p(epidural))* (1- p(subdural)) * ....</p>\n\n<p>```</p>\n\n<p>hence 6 class sigmoid loss is not the best</p>",
  "messages": [
    {
      "id": 661262,
      "postDate": "2019-10-30T04:10:06.473Z",
      "content": "<p>The problem with that formula is that the classes are correlated, in some non-trivial way. The independence assumption that you made is incorrect. It is even not clear if there are any negative correlations on some specific subgroups. </p>",
      "rawMarkdown": "The problem with that formula is that the classes are correlated, in some non-trivial way. The independence assumption that you made is incorrect. It is even not clear if there are any negative correlations on some specific subgroups. ",
      "votes": 5,
      "replies": [
        {
          "id": 661328,
          "postDate": "2019-10-30T06:28:43.293Z",
          "content": "<p>you are correct. that is why directly training 5 class don't work.</p>\n\n<p>\"... classes are correlated\" i will try a lstm layer to refine the labels</p>\n\n<p>thanks!</p>",
          "rawMarkdown": "you are correct. that is why directly training 5 class don't work.\n\n\"... classes are correlated\" i will try a lstm layer to refine the labels\n\nthanks!",
          "votes": 1
        }
      ]
    },
    {
      "id": 660868,
      "postDate": "2019-10-29T18:00:25.063Z",
      "content": "<p>shouldn't be it be  : \n```\np(any)  = 1 - p(not any) = 1 - (1-p(epidural))* (1- p(subdural)) * ....</p>\n\n<p>```</p>\n\n<p>hence 6 class sigmoid loss is not the best</p>",
      "rawMarkdown": "shouldn't be it be  : \n```\np(any)  = 1 - p(not any) = 1 - (1-p(epidural))* (1- p(subdural)) * ....\n\n```\n\nhence 6 class sigmoid loss is not the best",
      "votes": 5
    },
    {
      "id": 660967,
      "postDate": "2019-10-29T20:19:24.870Z",
      "content": "<p>Tried predicting one of 5 classes and calculating any. It gave substantially worse result. By the way, shouldn't it be summation instead of multiplication if dealing with 'one of many' probability ? </p>",
      "rawMarkdown": "Tried predicting one of 5 classes and calculating any. It gave substantially worse result. By the way, shouldn't it be summation instead of multiplication if dealing with 'one of many' probability ? ",
      "votes": 1,
      "replies": [
        {
          "id": 661246,
          "postDate": "2019-10-30T03:41:54.020Z",
          "content": "<p>direct 5 class prediction won't work. </p>\n\n<p>if you analyse your results, you will find that the majority of the score contribution comes from correct prediction of the negative samples.</p>\n\n<p>negative precision is better than the positive precision of the 5 class. hence to deduce the value of \"any\" from the five class is worse than to predict \"any\"  directly.</p>\n\n<p>now since one is better than the other and the two quantities are related, we should use the better value to correct the worse one.</p>\n\n<p>it is like:</p>\n\n<p>something else (better) &gt; 6-class sigmoid  &gt; 5-class sigmoid</p>",
          "rawMarkdown": "direct 5 class prediction won't work. \n\nif you analyse your results, you will find that the majority of the score contribution comes from correct prediction of the negative samples.\n\nnegative precision is better than the positive precision of the 5 class. hence to deduce the value of \"any\" from the five class is worse than to predict \"any\"  directly.\n\nnow since one is better than the other and the two quantities are related, we should use the better value to correct the worse one.\n\n it is like:\n\nsomething else (better) &gt; 6-class sigmoid  &gt; 5-class sigmoid"
        }
      ]
    },
    {
      "id": 662069,
      "postDate": "2019-10-31T04:04:04.323Z",
      "content": "<p><a href=\"/hengck23\">@hengck23</a> \ncan you please explain in more detail what u r trying to do \" i will try a lstm layer to refine the labels\"</p>",
      "rawMarkdown": "@hengck23 \ncan you please explain in more detail what u r trying to do \" i will try a lstm layer to refine the labels\"",
      "replies": [
        {
          "id": 662481,
          "postDate": "2019-10-31T15:50:00.997Z",
          "content": "<p>refer to <a href=\"https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109272#latest-661699\">https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109272#latest-661699</a></p>",
          "rawMarkdown": "refer to https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109272#latest-661699"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 661262,
      "author_name": "nosound",
      "author_url": "",
      "post_date": "2019-10-30T04:10:06.473000",
      "content": "<p>The problem with that formula is that the classes are correlated, in some non-trivial way. The independence assumption that you made is incorrect. It is even not clear if there are any negative correlations on some specific subgroups. </p>",
      "votes": 5,
      "replies": [
        {
          "id": 661328,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2019-10-30T06:28:43.293000",
          "content": "<p>you are correct. that is why directly training 5 class don't work.</p>\n\n<p>\"... classes are correlated\" i will try a lstm layer to refine the labels</p>\n\n<p>thanks!</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 660967,
      "author_name": "Eek The Cat",
      "author_url": "",
      "post_date": "2019-10-29T20:19:24.870000",
      "content": "<p>Tried predicting one of 5 classes and calculating any. It gave substantially worse result. By the way, shouldn't it be summation instead of multiplication if dealing with 'one of many' probability ? </p>",
      "votes": 1,
      "replies": [
        {
          "id": 661246,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2019-10-30T03:41:54.020000",
          "content": "<p>direct 5 class prediction won't work. </p>\n\n<p>if you analyse your results, you will find that the majority of the score contribution comes from correct prediction of the negative samples.</p>\n\n<p>negative precision is better than the positive precision of the 5 class. hence to deduce the value of \"any\" from the five class is worse than to predict \"any\"  directly.</p>\n\n<p>now since one is better than the other and the two quantities are related, we should use the better value to correct the worse one.</p>\n\n<p>it is like:</p>\n\n<p>something else (better) &gt; 6-class sigmoid  &gt; 5-class sigmoid</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 662069,
      "author_name": "NeerajSharma",
      "author_url": "",
      "post_date": "2019-10-31T04:04:04.323000",
      "content": "<p><a href=\"/hengck23\">@hengck23</a> \ncan you please explain in more detail what u r trying to do \" i will try a lstm layer to refine the labels\"</p>",
      "votes": 0,
      "replies": [
        {
          "id": 662481,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "2019-10-31T15:50:00.997000",
          "content": "<p>refer to <a href=\"https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109272#latest-661699\">https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109272#latest-661699</a></p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "661262": "The problem with that formula is that the classes are correlated, in some non-trivial way. The independence assumption that you made is incorrect. It is even not clear if there are any negative correlations on some specific subgroups. ",
    "660868": "shouldn't be it be  : \n```\np(any)  = 1 - p(not any) = 1 - (1-p(epidural))* (1- p(subdural)) * ....\n\n```\n\nhence 6 class sigmoid loss is not the best",
    "660967": "Tried predicting one of 5 classes and calculating any. It gave substantially worse result. By the way, shouldn't it be summation instead of multiplication if dealing with 'one of many' probability ? ",
    "662069": "@hengck23 \ncan you please explain in more detail what u r trying to do \" i will try a lstm layer to refine the labels\""
  }
}