{
  "id": 99663,
  "title": "Why better quadratic weighted kappa and better accuracy obtained a worse result",
  "url": "/competitions/aptos2019-blindness-detection/discussion/99663",
  "author_name": "",
  "post_date": "2019-07-13T05:39:20.452501400Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I train the model with the 90% train data and verify with the 10% train data, My model is se_resnext101 </p>\n\nmodel1 (regression):\n\n<p>cv: quadratic weighted kappa ~ 0.91,  accuracy_score ~ 0.80,  LB: 0.634</p>\n\nmodel2 (classification):\n\n<p>cv: quadratic weighted kappa ~ 0.92,  accuracy_score ~ 0.88,  LB: 0.500\nI think better accuracy and kappa coefficient will get better scores，but it turned out I was wrong. I am new to image classification, can anyone give me some Suggestions. Thank you very much</p>",
  "messages": [
    {
      "id": "574010",
      "postDate": "07/13/2019 05:39:20",
      "content": "<p>I train the model with the 90% train data and verify with the 10% train data, My model is se_resnext101 </p>\n\nmodel1 (regression):\n\n<p>cv: quadratic weighted kappa ~ 0.91,  accuracy_score ~ 0.80,  LB: 0.634</p>\n\nmodel2 (classification):\n\n<p>cv: quadratic weighted kappa ~ 0.92,  accuracy_score ~ 0.88,  LB: 0.500\nI think better accuracy and kappa coefficient will get better scores，but it turned out I was wrong. I am new to image classification, can anyone give me some Suggestions. Thank you very much</p>",
      "rawMarkdown": "I train the model with the 90% train data and verify with the 10% train data, My model is se_resnext101 \n###### model1 (regression):\ncv: quadratic weighted kappa ~ 0.91,  accuracy_score ~ 0.80,  LB: 0.634\n###### model2 (classification):\ncv: quadratic weighted kappa ~ 0.92,  accuracy_score ~ 0.88,  LB: 0.500\nI think better accuracy and kappa coefficient will get better scores，but it turned out I was wrong. I am new to image classification, can anyone give me some Suggestions. Thank you very much",
      "votes": null
    },
    {
      "id": "574122",
      "postDate": "07/13/2019 09:14:47",
      "content": "<p>Because   the   distribution     in    LB   dataset   is   different  from   your  trainset  ,  your   model2 (classification)   overfit   to  your   train_dataset.\nOr  you  can  have   a  try   in    K-Fold .</p>",
      "rawMarkdown": "Because   the   distribution     in    LB   dataset   is   different  from   your  trainset  ,  your   model2 (classification)   overfit   to  your   train_dataset.\nOr  you  can  have   a  try   in    K-Fold .",
      "votes": null
    },
    {
      "id": "576915",
      "postDate": "07/16/2019 07:00:15",
      "content": "<p>Thank you for your suggestions, I will try more for it!!!!</p>",
      "rawMarkdown": "Thank you for your suggestions, I will try more for it!!!!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 574122,
      "author_name": "xujingzhao",
      "author_url": "",
      "post_date": "07/13/2019 09:14:47",
      "content": "<p>Because   the   distribution     in    LB   dataset   is   different  from   your  trainset  ,  your   model2 (classification)   overfit   to  your   train_dataset.\nOr  you  can  have   a  try   in    K-Fold .</p>",
      "votes": null,
      "replies": [
        {
          "id": 576915,
          "author_name": "hesene",
          "author_url": "",
          "post_date": "07/16/2019 07:00:15",
          "content": "<p>Thank you for your suggestions, I will try more for it!!!!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "574010": "I train the model with the 90% train data and verify with the 10% train data, My model is se_resnext101 \n###### model1 (regression):\ncv: quadratic weighted kappa ~ 0.91,  accuracy_score ~ 0.80,  LB: 0.634\n###### model2 (classification):\ncv: quadratic weighted kappa ~ 0.92,  accuracy_score ~ 0.88,  LB: 0.500\nI think better accuracy and kappa coefficient will get better scores，but it turned out I was wrong. I am new to image classification, can anyone give me some Suggestions. Thank you very much",
    "574122": "Because   the   distribution     in    LB   dataset   is   different  from   your  trainset  ,  your   model2 (classification)   overfit   to  your   train_dataset.\nOr  you  can  have   a  try   in    K-Fold .",
    "576915": "Thank you for your suggestions, I will try more for it!!!!"
  },
  "source": "meta"
}