{
  "id": 81839,
  "title": "Which customers are correlated in the training set for proper cv?",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/81839",
  "author_name": "",
  "post_date": "2019-02-25T14:33:28.169385600Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I have a feeling that the customers  in training data have high correlation. Originally I think the target score have lognormal distribution which is typical for correlated data. If this is the case then that explains the difference in CV and LB scores.</p>",
  "messages": [
    {
      "id": "477940",
      "postDate": "02/25/2019 14:33:28",
      "content": "<p>I have a feeling that the customers  in training data have high correlation. Originally I think the target score have lognormal distribution which is typical for correlated data. If this is the case then that explains the difference in CV and LB scores.</p>",
      "rawMarkdown": "I have a feeling that the customers  in training data have high correlation. Originally I think the target score have lognormal distribution which is typical for correlated data. If this is the case then that explains the difference in CV and LB scores.",
      "votes": null
    },
    {
      "id": "477942",
      "postDate": "02/25/2019 14:37:53",
      "content": "<p>yes, target is lognormal distributed.</p>",
      "rawMarkdown": "yes, target is lognormal distributed.",
      "votes": null
    },
    {
      "id": "477967",
      "postDate": "02/25/2019 15:01:44",
      "content": "<p>But I kinda changing mind. If you do prover CV without thinking about correlated customers, LB and CV will not vary that much. </p>",
      "rawMarkdown": "But I kinda changing mind. If you do prover CV without thinking about correlated customers, LB and CV will not vary that much.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 477942,
      "author_name": "liewmanchoi",
      "author_url": "",
      "post_date": "02/25/2019 14:37:53",
      "content": "<p>yes, target is lognormal distributed.</p>",
      "votes": null,
      "replies": [
        {
          "id": 477967,
          "author_name": "aceplayer11",
          "author_url": "",
          "post_date": "02/25/2019 15:01:44",
          "content": "<p>But I kinda changing mind. If you do prover CV without thinking about correlated customers, LB and CV will not vary that much. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "477940": "I have a feeling that the customers  in training data have high correlation. Originally I think the target score have lognormal distribution which is typical for correlated data. If this is the case then that explains the difference in CV and LB scores.",
    "477942": "yes, target is lognormal distributed.",
    "477967": "But I kinda changing mind. If you do prover CV without thinking about correlated customers, LB and CV will not vary that much."
  },
  "source": "meta"
}