{
  "id": 55410,
  "title": "Why my local CV is a bit higher than locan train-set CV?",
  "url": "/competitions/talkingdata-adtracking-fraud-detection/discussion/55410",
  "author_name": "",
  "post_date": "2018-04-26T08:53:38.763499200Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Is it because the data has time? I use train-test-split to get train and validation data.. Maybe this is wrong... </p>",
  "messages": [
    {
      "id": "319536",
      "postDate": "04/26/2018 08:53:38",
      "content": "<p>Is it because the data has time? I use train-test-split to get train and validation data.. Maybe this is wrong... </p>",
      "rawMarkdown": "Is it because the data has time? I use train-test-split to get train and validation data.. Maybe this is wrong...",
      "votes": null
    },
    {
      "id": "319559",
      "postDate": "04/26/2018 10:01:46",
      "content": "<p>Random train/test?\nIt has data leakage from the future that you don't have in the real test data.</p>\n\n<p>At least use the last samples of the training.\nOr better, use the first two days for training and simulate the real test on the last day and use the real hours used in it. There is quite a big difference in AUC in the different hours</p>",
      "rawMarkdown": "Random train/test?\nIt has data leakage from the future that you don't have in the real test data.\n\nAt least use the last samples of the training.\nOr better, use the first two days for training and simulate the real test on the last day and use the real hours used in it. There is quite a big difference in AUC in the different hours",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 319559,
      "author_name": "mrbeer",
      "author_url": "",
      "post_date": "04/26/2018 10:01:46",
      "content": "<p>Random train/test?\nIt has data leakage from the future that you don't have in the real test data.</p>\n\n<p>At least use the last samples of the training.\nOr better, use the first two days for training and simulate the real test on the last day and use the real hours used in it. There is quite a big difference in AUC in the different hours</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "319536": "Is it because the data has time? I use train-test-split to get train and validation data.. Maybe this is wrong...",
    "319559": "Random train/test?\nIt has data leakage from the future that you don't have in the real test data.\n\nAt least use the last samples of the training.\nOr better, use the first two days for training and simulate the real test on the last day and use the real hours used in it. There is quite a big difference in AUC in the different hours"
  },
  "source": "meta"
}