{
  "id": 164617,
  "title": "Is predicting Diagnosis better than predicting Target?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/164617",
  "author_name": "",
  "post_date": "2020-07-06T23:02:14.871800800Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Because the metric of this model is 'AUC' - so to get a high score we need the model to predict higher probability for images with malign melonama. </p>\n\n<p>Would predicting diagnosis instead of Target in the train.csv file help improve performance? We could then take out all the predictions made for MELONAMA category, do a sigmoid and submit those instead.</p>\n\n<p>Is this the secret sauce to get an even higher score? Has anyone tried this?</p>",
  "messages": [
    {
      "id": "918042",
      "postDate": "07/06/2020 23:02:14",
      "content": "<p>Because the metric of this model is 'AUC' - so to get a high score we need the model to predict higher probability for images with malign melonama. </p>\n\n<p>Would predicting diagnosis instead of Target in the train.csv file help improve performance? We could then take out all the predictions made for MELONAMA category, do a sigmoid and submit those instead.</p>\n\n<p>Is this the secret sauce to get an even higher score? Has anyone tried this?</p>",
      "rawMarkdown": "Because the metric of this model is 'AUC' - so to get a high score we need the model to predict higher probability for images with malign melonama. \n\nWould predicting diagnosis instead of Target in the train.csv file help improve performance? We could then take out all the predictions made for MELONAMA category, do a sigmoid and submit those instead.\n\nIs this the secret sauce to get an even higher score? Has anyone tried this?",
      "votes": null
    },
    {
      "id": "918108",
      "postDate": "07/07/2020 02:00:08",
      "content": "<p>Diagnosis column has too many unknowns:\n-       1  atypical melanocytic proliferation\n-          1  cafe-au-lait macule\n-      44  lentigo NOS\n-      37  lichenoid keratosis\n-     584  melanoma\n-    5193  nevus\n-     135  seborrheic keratosis\n-       7  solar lentigo\n-   27124  unknown</p>\n\n<p>This kind of analysis may not help in identifying Ms directly, but may be more useful in removing False Positives such as Seborrheic Keratosis being identified incorrectly as M.</p>",
      "rawMarkdown": "Diagnosis column has too many unknowns:\n-       1  atypical melanocytic proliferation\n-          1  cafe-au-lait macule\n-      44  lentigo NOS\n-      37  lichenoid keratosis\n-     584  melanoma\n-    5193  nevus\n-     135  seborrheic keratosis\n-       7  solar lentigo\n-   27124  unknown\n\nThis kind of analysis may not help in identifying Ms directly, but may be more useful in removing False Positives such as Seborrheic Keratosis being identified incorrectly as M.",
      "votes": null
    },
    {
      "id": "918173",
      "postDate": "07/07/2020 03:42:41",
      "content": "<p>Could I please ask how could we use it to remove False Positives? Also, I believe having False Positives wouldn't hurt the score much.</p>\n\n<p>It's the False Negatives that hurt it the most. </p>",
      "rawMarkdown": "Could I please ask how could we use it to remove False Positives? Also, I believe having False Positives wouldn't hurt the score much.\n\nIt's the False Negatives that hurt it the most.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 918108,
      "author_name": "sirishks",
      "author_url": "",
      "post_date": "07/07/2020 02:00:08",
      "content": "<p>Diagnosis column has too many unknowns:\n-       1  atypical melanocytic proliferation\n-          1  cafe-au-lait macule\n-      44  lentigo NOS\n-      37  lichenoid keratosis\n-     584  melanoma\n-    5193  nevus\n-     135  seborrheic keratosis\n-       7  solar lentigo\n-   27124  unknown</p>\n\n<p>This kind of analysis may not help in identifying Ms directly, but may be more useful in removing False Positives such as Seborrheic Keratosis being identified incorrectly as M.</p>",
      "votes": null,
      "replies": [
        {
          "id": 918173,
          "author_name": "aroraaman",
          "author_url": "",
          "post_date": "07/07/2020 03:42:41",
          "content": "<p>Could I please ask how could we use it to remove False Positives? Also, I believe having False Positives wouldn't hurt the score much.</p>\n\n<p>It's the False Negatives that hurt it the most. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "918042": "Because the metric of this model is 'AUC' - so to get a high score we need the model to predict higher probability for images with malign melonama. \n\nWould predicting diagnosis instead of Target in the train.csv file help improve performance? We could then take out all the predictions made for MELONAMA category, do a sigmoid and submit those instead.\n\nIs this the secret sauce to get an even higher score? Has anyone tried this?",
    "918108": "Diagnosis column has too many unknowns:\n-       1  atypical melanocytic proliferation\n-          1  cafe-au-lait macule\n-      44  lentigo NOS\n-      37  lichenoid keratosis\n-     584  melanoma\n-    5193  nevus\n-     135  seborrheic keratosis\n-       7  solar lentigo\n-   27124  unknown\n\nThis kind of analysis may not help in identifying Ms directly, but may be more useful in removing False Positives such as Seborrheic Keratosis being identified incorrectly as M.",
    "918173": "Could I please ask how could we use it to remove False Positives? Also, I believe having False Positives wouldn't hurt the score much.\n\nIt's the False Negatives that hurt it the most."
  },
  "source": "meta"
}