{
  "id": 103071,
  "title": "Why so much variation ??",
  "url": "/competitions/aptos2019-blindness-detection/discussion/103071",
  "author_name": "",
  "post_date": "2019-08-06T23:06:43.733373700Z",
  "votes": null,
  "comment_count": 10,
  "views": 0,
  "content": "<p>I am getting accuracy of 80%+ but kappa value in negative..\ncan anybody help me..</p>",
  "messages": [
    {
      "id": "593645",
      "postDate": "08/06/2019 23:06:43",
      "content": "<p>I am getting accuracy of 80%+ but kappa value in negative..\ncan anybody help me..</p>",
      "rawMarkdown": "I am getting accuracy of 80%+ but kappa value in negative..\ncan anybody help me..",
      "votes": null
    },
    {
      "id": "593781",
      "postDate": "08/07/2019 05:25:24",
      "content": "<p>Refer to this kernel: <a href=\"https://www.kaggle.com/taindow/be-careful-what-you-train-on\">https://www.kaggle.com/taindow/be-careful-what-you-train-on</a>\nMaybe your model is learning bias like width/height ratio of an image.</p>",
      "rawMarkdown": "Refer to this kernel: https://www.kaggle.com/taindow/be-careful-what-you-train-on\nMaybe your model is learning bias like width/height ratio of an image.",
      "votes": null
    },
    {
      "id": "594040",
      "postDate": "08/07/2019 13:34:56",
      "content": "<p>If you predict only 2s, than you get 80% accuracy in the public test set, but it's obviously a bad aproach.</p>",
      "rawMarkdown": "If you predict only 2s, than you get 80% accuracy in the public test set, but it's obviously a bad aproach.",
      "votes": null
    },
    {
      "id": "594403",
      "postDate": "08/08/2019 00:49:29",
      "content": "<p>How do you know the accuracy of your predictions in the public test set?</p>",
      "rawMarkdown": "How do you know the accuracy of your predictions in the public test set?",
      "votes": null
    },
    {
      "id": "594540",
      "postDate": "08/08/2019 06:36:48",
      "content": "<p>LB score is measured on public test set only :)</p>",
      "rawMarkdown": "LB score is measured on public test set only :)",
      "votes": null
    },
    {
      "id": "594541",
      "postDate": "08/08/2019 06:38:08",
      "content": "<p>Hey <a href=\"/niteshfre\">@niteshfre</a> There are a few threads in the forum where people have discussed this issue, find them out :)</p>",
      "rawMarkdown": "Hey @niteshfre There are a few threads in the forum where people have discussed this issue, find them out :)",
      "votes": null
    },
    {
      "id": "594553",
      "postDate": "08/08/2019 06:49:04",
      "content": "<p>I mean, we can only know quadratic kappa score, don't we? If I submit all 2s, my public lb score is 0.</p>",
      "rawMarkdown": "I mean, we can only know quadratic kappa score, don't we? If I submit all 2s, my public lb score is 0.",
      "votes": null
    },
    {
      "id": "594660",
      "postDate": "08/08/2019 08:54:42",
      "content": "<p>thank you</p>",
      "rawMarkdown": "thank you",
      "votes": null
    },
    {
      "id": "594662",
      "postDate": "08/08/2019 08:55:30",
      "content": "<p>thanks, for giving new vision think..</p>",
      "rawMarkdown": "thanks, for giving new vision think..",
      "votes": null
    },
    {
      "id": "594810",
      "postDate": "08/08/2019 13:23:02",
      "content": "<ol>\n<li>It would be better to specify on what data you get which  result - do you mean 80% accuracy on holdout(CV) or train and negative LB, or both 80% acc and negative cappa on the same data(holdout/CV)?</li>\n<li>As people already mentioned, look at you confusion matrix or per class accuracy how well do you do on each class.</li>\n<li>If you have good CV but bad LB - look at the predicted class distribution in your submission.csv. Does it look meaningful?</li>\n</ol>",
      "rawMarkdown": "1. It would be better to specify on what data you get which  result - do you mean 80% accuracy on holdout(CV) or train and negative LB, or both 80% acc and negative cappa on the same data(holdout/CV)?\n2. As people already mentioned, look at you confusion matrix or per class accuracy how well do you do on each class.\n3. If you have good CV but bad LB - look at the predicted class distribution in your submission.csv. Does it look meaningful?",
      "votes": null
    },
    {
      "id": "595211",
      "postDate": "08/09/2019 00:29:06",
      "content": "<p>I assumed he looked at his best submission and saw the proportion of 2s</p>",
      "rawMarkdown": "I assumed he looked at his best submission and saw the proportion of 2s",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 593781,
      "author_name": "harangdev",
      "author_url": "",
      "post_date": "08/07/2019 05:25:24",
      "content": "<p>Refer to this kernel: <a href=\"https://www.kaggle.com/taindow/be-careful-what-you-train-on\">https://www.kaggle.com/taindow/be-careful-what-you-train-on</a>\nMaybe your model is learning bias like width/height ratio of an image.</p>",
      "votes": null,
      "replies": [
        {
          "id": 594662,
          "author_name": "niteshfre",
          "author_url": "",
          "post_date": "08/08/2019 08:55:30",
          "content": "<p>thanks, for giving new vision think..</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 594040,
      "author_name": "nemethpeti",
      "author_url": "",
      "post_date": "08/07/2019 13:34:56",
      "content": "<p>If you predict only 2s, than you get 80% accuracy in the public test set, but it's obviously a bad aproach.</p>",
      "votes": null,
      "replies": [
        {
          "id": 594403,
          "author_name": "harangdev",
          "author_url": "",
          "post_date": "08/08/2019 00:49:29",
          "content": "<p>How do you know the accuracy of your predictions in the public test set?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 594540,
          "author_name": "rishabhiitbhu",
          "author_url": "",
          "post_date": "08/08/2019 06:36:48",
          "content": "<p>LB score is measured on public test set only :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 594553,
          "author_name": "harangdev",
          "author_url": "",
          "post_date": "08/08/2019 06:49:04",
          "content": "<p>I mean, we can only know quadratic kappa score, don't we? If I submit all 2s, my public lb score is 0.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 595211,
          "author_name": "sdoria",
          "author_url": "",
          "post_date": "08/09/2019 00:29:06",
          "content": "<p>I assumed he looked at his best submission and saw the proportion of 2s</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 594541,
      "author_name": "rishabhiitbhu",
      "author_url": "",
      "post_date": "08/08/2019 06:38:08",
      "content": "<p>Hey <a href=\"/niteshfre\">@niteshfre</a> There are a few threads in the forum where people have discussed this issue, find them out :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 594660,
          "author_name": "niteshfre",
          "author_url": "",
          "post_date": "08/08/2019 08:54:42",
          "content": "<p>thank you</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 594810,
      "author_name": "vzaguskin",
      "author_url": "",
      "post_date": "08/08/2019 13:23:02",
      "content": "<ol>\n<li>It would be better to specify on what data you get which  result - do you mean 80% accuracy on holdout(CV) or train and negative LB, or both 80% acc and negative cappa on the same data(holdout/CV)?</li>\n<li>As people already mentioned, look at you confusion matrix or per class accuracy how well do you do on each class.</li>\n<li>If you have good CV but bad LB - look at the predicted class distribution in your submission.csv. Does it look meaningful?</li>\n</ol>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "593645": "I am getting accuracy of 80%+ but kappa value in negative..\ncan anybody help me..",
    "593781": "Refer to this kernel: https://www.kaggle.com/taindow/be-careful-what-you-train-on\nMaybe your model is learning bias like width/height ratio of an image.",
    "594040": "If you predict only 2s, than you get 80% accuracy in the public test set, but it's obviously a bad aproach.",
    "594403": "How do you know the accuracy of your predictions in the public test set?",
    "594540": "LB score is measured on public test set only :)",
    "594541": "Hey @niteshfre There are a few threads in the forum where people have discussed this issue, find them out :)",
    "594553": "I mean, we can only know quadratic kappa score, don't we? If I submit all 2s, my public lb score is 0.",
    "594660": "thank you",
    "594662": "thanks, for giving new vision think..",
    "594810": "1. It would be better to specify on what data you get which  result - do you mean 80% accuracy on holdout(CV) or train and negative LB, or both 80% acc and negative cappa on the same data(holdout/CV)?\n2. As people already mentioned, look at you confusion matrix or per class accuracy how well do you do on each class.\n3. If you have good CV but bad LB - look at the predicted class distribution in your submission.csv. Does it look meaningful?",
    "595211": "I assumed he looked at his best submission and saw the proportion of 2s"
  },
  "source": "meta"
}