{
  "id": 84641,
  "title": "Understand more on MCC and perhaps devise a new strategy",
  "url": "/competitions/vsb-power-line-fault-detection/discussion/84641",
  "author_name": "",
  "post_date": "2019-03-18T15:54:42.429935200Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Greeting everyone. Let us try to get more insight from the MCC formula  : </p>\n\n<p>$$\n\\text{MCC} = \\frac{ \\mathit{TP} \\times \\mathit{TN} - \\mathit{FP} \\times \\mathit{FN} } {\\sqrt{ (\\mathit{TP} + \\mathit{FP}) ( \\mathit{TP} + \\mathit{FN} ) ( \\mathit{TN} + \\mathit{FP} ) ( \\mathit{TN} + \\mathit{FN} ) } }\n$$</p>\n\n<p>The formula is quite complex and not so intuitive compared to usual metrics like <strong>accuracy, F1</strong>, etc. Therefore, we will have a hard time to gain insight from our current MCC score, e.g. if we score 0.694, </p>\n\n<ul>\n<li><p>how many positive/negative examples that we are correctly classified? </p></li>\n<li><p>Does our algorithm have a good precision?</p></li>\n<li><p>If we are able to correct one more positive / negative example, how much will MCC increase?</p></li>\n</ul>\n\n<p>Even though we may not answer the above questions perfectly, in this problem, we may have a way to approximate them. I wrote one way to analyze our MCC score and answer the above questions in this kernel :</p>\n\n<p><a href=\"https://www.kaggle.com/ratthachat/a-heuristic-to-understand-your-mcc/\">https://www.kaggle.com/ratthachat/a-heuristic-to-understand-your-mcc/</a></p>\n\n<p>Hope it be a little helpful!</p>",
  "messages": [
    {
      "id": "493380",
      "postDate": "03/18/2019 15:54:42",
      "content": "<p>Greeting everyone. Let us try to get more insight from the MCC formula  : </p>\n\n<p>$$\n\\text{MCC} = \\frac{ \\mathit{TP} \\times \\mathit{TN} - \\mathit{FP} \\times \\mathit{FN} } {\\sqrt{ (\\mathit{TP} + \\mathit{FP}) ( \\mathit{TP} + \\mathit{FN} ) ( \\mathit{TN} + \\mathit{FP} ) ( \\mathit{TN} + \\mathit{FN} ) } }\n$$</p>\n\n<p>The formula is quite complex and not so intuitive compared to usual metrics like <strong>accuracy, F1</strong>, etc. Therefore, we will have a hard time to gain insight from our current MCC score, e.g. if we score 0.694, </p>\n\n<ul>\n<li><p>how many positive/negative examples that we are correctly classified? </p></li>\n<li><p>Does our algorithm have a good precision?</p></li>\n<li><p>If we are able to correct one more positive / negative example, how much will MCC increase?</p></li>\n</ul>\n\n<p>Even though we may not answer the above questions perfectly, in this problem, we may have a way to approximate them. I wrote one way to analyze our MCC score and answer the above questions in this kernel :</p>\n\n<p><a href=\"https://www.kaggle.com/ratthachat/a-heuristic-to-understand-your-mcc/\">https://www.kaggle.com/ratthachat/a-heuristic-to-understand-your-mcc/</a></p>\n\n<p>Hope it be a little helpful!</p>",
      "rawMarkdown": "Greeting everyone. Let us try to get more insight from the MCC formula  : \n\n$$\n\\text{MCC} = \\frac{ \\mathit{TP} \\times \\mathit{TN} - \\mathit{FP} \\times \\mathit{FN} } {\\sqrt{ (\\mathit{TP} + \\mathit{FP}) ( \\mathit{TP} + \\mathit{FN} ) ( \\mathit{TN} + \\mathit{FP} ) ( \\mathit{TN} + \\mathit{FN} ) } }\n$$\n\nThe formula is quite complex and not so intuitive compared to usual metrics like **accuracy, F1**, etc. Therefore, we will have a hard time to gain insight from our current MCC score, e.g. if we score 0.694, \n\n- how many positive/negative examples that we are correctly classified? \n\n- Does our algorithm have a good precision?\n\n- If we are able to correct one more positive / negative example, how much will MCC increase?\n\nEven though we may not answer the above questions perfectly, in this problem, we may have a way to approximate them. I wrote one way to analyze our MCC score and answer the above questions in this kernel :\n\nhttps://www.kaggle.com/ratthachat/a-heuristic-to-understand-your-mcc/\n\nHope it be a little helpful!",
      "votes": null
    },
    {
      "id": "493455",
      "postDate": "03/18/2019 16:49:36",
      "content": "<p>I like this kernel. Not much to read, but gives a great intuition <a href=\"https://www.kaggle.com/kenmatsu4/intuitive-interpretation-of-mcc\">https://www.kaggle.com/kenmatsu4/intuitive-interpretation-of-mcc</a></p>",
      "rawMarkdown": "I like this kernel. Not much to read, but gives a great intuition https://www.kaggle.com/kenmatsu4/intuitive-interpretation-of-mcc",
      "votes": null
    },
    {
      "id": "493722",
      "postDate": "03/19/2019 02:24:33",
      "content": "<p>Hi DavidS, </p>\n\n<p>Yes! I like that kernel as well. I used it  as a starting point when I entered the competition.</p>",
      "rawMarkdown": "Hi DavidS, \n\nYes! I like that kernel as well. I used it  as a starting point when I entered the competition.",
      "votes": null
    },
    {
      "id": "493774",
      "postDate": "03/19/2019 04:37:55",
      "content": "<p>Great work <a href=\"/ratthachat\">@ratthachat</a> ! This is exactly what I need now !</p>",
      "rawMarkdown": "Great work @ratthachat ! This is exactly what I need now !",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 493455,
      "author_name": "davids1992",
      "author_url": "",
      "post_date": "03/18/2019 16:49:36",
      "content": "<p>I like this kernel. Not much to read, but gives a great intuition <a href=\"https://www.kaggle.com/kenmatsu4/intuitive-interpretation-of-mcc\">https://www.kaggle.com/kenmatsu4/intuitive-interpretation-of-mcc</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 493722,
          "author_name": "ratthachat",
          "author_url": "",
          "post_date": "03/19/2019 02:24:33",
          "content": "<p>Hi DavidS, </p>\n\n<p>Yes! I like that kernel as well. I used it  as a starting point when I entered the competition.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 493774,
      "author_name": "tarunpaparaju",
      "author_url": "",
      "post_date": "03/19/2019 04:37:55",
      "content": "<p>Great work <a href=\"/ratthachat\">@ratthachat</a> ! This is exactly what I need now !</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "493380": "Greeting everyone. Let us try to get more insight from the MCC formula  : \n\n$$\n\\text{MCC} = \\frac{ \\mathit{TP} \\times \\mathit{TN} - \\mathit{FP} \\times \\mathit{FN} } {\\sqrt{ (\\mathit{TP} + \\mathit{FP}) ( \\mathit{TP} + \\mathit{FN} ) ( \\mathit{TN} + \\mathit{FP} ) ( \\mathit{TN} + \\mathit{FN} ) } }\n$$\n\nThe formula is quite complex and not so intuitive compared to usual metrics like **accuracy, F1**, etc. Therefore, we will have a hard time to gain insight from our current MCC score, e.g. if we score 0.694, \n\n- how many positive/negative examples that we are correctly classified? \n\n- Does our algorithm have a good precision?\n\n- If we are able to correct one more positive / negative example, how much will MCC increase?\n\nEven though we may not answer the above questions perfectly, in this problem, we may have a way to approximate them. I wrote one way to analyze our MCC score and answer the above questions in this kernel :\n\nhttps://www.kaggle.com/ratthachat/a-heuristic-to-understand-your-mcc/\n\nHope it be a little helpful!",
    "493455": "I like this kernel. Not much to read, but gives a great intuition https://www.kaggle.com/kenmatsu4/intuitive-interpretation-of-mcc",
    "493722": "Hi DavidS, \n\nYes! I like that kernel as well. I used it  as a starting point when I entered the competition.",
    "493774": "Great work @ratthachat ! This is exactly what I need now !"
  },
  "source": "meta"
}