{
  "id": 82073,
  "title": "seed 2029 !!! maybe this is where pub private spilts",
  "url": "/competitions/elo-merchant-category-recommendation/discussion/82073",
  "author_name": "",
  "post_date": "2019-02-27T04:22:24.098146Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>local cv with seed 2029 Kfold( 10 )  gives 3 folds over 3.75 and 7 folds under 3.62, so I suspect this is where private/LB divided(  something like  seed 2029 split 10 Kfold  by target  3 folds for public and 3 folds for private)  !!!! any thoughs!!!</p>",
  "messages": [
    {
      "id": "479328",
      "postDate": "02/27/2019 04:22:24",
      "content": "<p>local cv with seed 2029 Kfold( 10 )  gives 3 folds over 3.75 and 7 folds under 3.62, so I suspect this is where private/LB divided(  something like  seed 2029 split 10 Kfold  by target  3 folds for public and 3 folds for private)  !!!! any thoughs!!!</p>",
      "rawMarkdown": "local cv with seed 2029 Kfold( 10 )  gives 3 folds over 3.75 and 7 folds under 3.62, so I suspect this is where private/LB divided(  something like  seed 2029 split 10 Kfold  by target  3 folds for public and 3 folds for private)  !!!! any thoughs!!!",
      "votes": null
    },
    {
      "id": "479689",
      "postDate": "02/27/2019 10:44:41",
      "content": "<p>?? test data does not overlap with train data.</p>",
      "rawMarkdown": "?? test data does not overlap with train data.",
      "votes": null
    },
    {
      "id": "479814",
      "postDate": "02/27/2019 12:54:46",
      "content": "<p>suspect that test has same distribution as train; we trained train_data get this outcome: 3 folds lower cv, and  7 higher which meet the pecentage of pub/private(  3 foldes has same distribute with public LB and 7 with private LB )  ,didn't you find that interesting? \n    may be through this we can fit beter for predit private LB, eg fit only use 7folds dataset</p>",
      "rawMarkdown": "suspect that test has same distribution as train; we trained train_data get this outcome: 3 folds lower cv, and  7 higher which meet the pecentage of pub/private(  3 foldes has same distribute with public LB and 7 with private LB )  ,didn't you find that interesting? \n    may be through this we can fit beter for predit private LB, eg fit only use 7folds dataset",
      "votes": null
    },
    {
      "id": "479909",
      "postDate": "02/27/2019 14:58:41",
      "content": "<p>There is a very easy way to test your hypothesis: use only the 7 higher folds and see is the test prediction is better.</p>",
      "rawMarkdown": "There is a very easy way to test your hypothesis: use only the 7 higher folds and see is the test prediction is better.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 479689,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "02/27/2019 10:44:41",
      "content": "<p>?? test data does not overlap with train data.</p>",
      "votes": null,
      "replies": [
        {
          "id": 479814,
          "author_name": "youwbpyr",
          "author_url": "",
          "post_date": "02/27/2019 12:54:46",
          "content": "<p>suspect that test has same distribution as train; we trained train_data get this outcome: 3 folds lower cv, and  7 higher which meet the pecentage of pub/private(  3 foldes has same distribute with public LB and 7 with private LB )  ,didn't you find that interesting? \n    may be through this we can fit beter for predit private LB, eg fit only use 7folds dataset</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 479909,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "02/27/2019 14:58:41",
          "content": "<p>There is a very easy way to test your hypothesis: use only the 7 higher folds and see is the test prediction is better.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "479328": "local cv with seed 2029 Kfold( 10 )  gives 3 folds over 3.75 and 7 folds under 3.62, so I suspect this is where private/LB divided(  something like  seed 2029 split 10 Kfold  by target  3 folds for public and 3 folds for private)  !!!! any thoughs!!!",
    "479689": "?? test data does not overlap with train data.",
    "479814": "suspect that test has same distribution as train; we trained train_data get this outcome: 3 folds lower cv, and  7 higher which meet the pecentage of pub/private(  3 foldes has same distribute with public LB and 7 with private LB )  ,didn't you find that interesting? \n    may be through this we can fit beter for predit private LB, eg fit only use 7folds dataset",
    "479909": "There is a very easy way to test your hypothesis: use only the 7 higher folds and see is the test prediction is better."
  },
  "source": "meta"
}