{
  "id": 287379,
  "title": "Weird right ?",
  "url": "/competitions/g-research-crypto-forecasting/discussion/287379",
  "author_name": "",
  "post_date": "2021-11-13T15:14:42.321468100Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi ,</p>\n<p>I have been into kaggle and machine learning for more than a year but i have never seen something like this thanks to the targets of the public lb present in train.csv . I noticed that most of the xgb and lgdm kernels were not doing something very important and common what say obvious people dont split between a validation and a train set like why ? I think as the targets of public lb are there in  <code>train.csv</code> they dont care about overfit and dont split the data and just overfit to the targets of public LB or overfitting to public lb in general . I think this is the reason though is foolish as they will just fail in private lb i want to listen to other peoples thoughts on this</p>\n<p>Thank you</p>",
  "messages": [
    {
      "id": "1581267",
      "postDate": "11/13/2021 15:14:42",
      "content": "<p>Hi ,</p>\n<p>I have been into kaggle and machine learning for more than a year but i have never seen something like this thanks to the targets of the public lb present in train.csv . I noticed that most of the xgb and lgdm kernels were not doing something very important and common what say obvious people dont split between a validation and a train set like why ? I think as the targets of public lb are there in  <code>train.csv</code> they dont care about overfit and dont split the data and just overfit to the targets of public LB or overfitting to public lb in general . I think this is the reason though is foolish as they will just fail in private lb i want to listen to other peoples thoughts on this</p>\n<p>Thank you</p>",
      "rawMarkdown": "Hi ,\n\nI have been into kaggle and machine learning for more than a year but i have never seen something like this thanks to the targets of the public lb present in train.csv . I noticed that most of the xgb and lgdm kernels were not doing something very important and common what say obvious people dont split between a validation and a train set like why ? I think as the targets of public lb are there in  ```train.csv``` they dont care about overfit and dont split the data and just overfit to the targets of public LB or overfitting to public lb in general . I think this is the reason though is foolish as they will just fail in private lb i want to listen to other peoples thoughts on this\n\nThank you",
      "votes": null
    },
    {
      "id": "1581426",
      "postDate": "11/13/2021 18:37:31",
      "content": "<p>I agree with you. However,  FYI you haven't excluded  the overlapping public LB samples from your kernel (LB 0.638) as well, so the score doesn't reflect the true generalisation performance of the model.  </p>",
      "rawMarkdown": "I agree with you. However,  FYI you haven't excluded  the overlapping public LB samples from your kernel (LB 0.638) as well, so the score doesn't reflect the true generalisation performance of the model.",
      "votes": null
    },
    {
      "id": "1582125",
      "postDate": "11/14/2021 15:05:31",
      "content": "<p>Hi,<br>\nit is always great to have some extra data from public lb or extra data in general i did not create that notebook for public lb extra data may improve overfit focus that i mean 'generalization' which means between Validation MSE and train MSE  i never said i wanted to generalize between public LB and CV . Other kernels are just overfitting on purpose as otherwise they would have done a train valid split to measure it i have reduced and simplified the LGDM model to fix overfit</p>",
      "rawMarkdown": "Hi,\nit is always great to have some extra data from public lb or extra data in general i did not create that notebook for public lb extra data may improve overfit focus that i mean 'generalization' which means between Validation MSE and train MSE  i never said i wanted to generalize between public LB and CV . Other kernels are just overfitting on purpose as otherwise they would have done a train valid split to measure it i have reduced and simplified the LGDM model to fix overfit",
      "votes": null
    },
    {
      "id": "1599871",
      "postDate": "11/29/2021 22:48:16",
      "content": "<p>90% of public notebooks made, to get additional votes as \"Notebooks Contributor\" and they don't have big benefits for users.</p>",
      "rawMarkdown": "90% of public notebooks made, to get additional votes as \"Notebooks Contributor\" and they don't have big benefits for users.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1581426,
      "author_name": "imeintanis",
      "author_url": "",
      "post_date": "11/13/2021 18:37:31",
      "content": "<p>I agree with you. However,  FYI you haven't excluded  the overlapping public LB samples from your kernel (LB 0.638) as well, so the score doesn't reflect the true generalisation performance of the model.  </p>",
      "votes": null,
      "replies": [
        {
          "id": 1582125,
          "author_name": "swaralipibose",
          "author_url": "",
          "post_date": "11/14/2021 15:05:31",
          "content": "<p>Hi,<br>\nit is always great to have some extra data from public lb or extra data in general i did not create that notebook for public lb extra data may improve overfit focus that i mean 'generalization' which means between Validation MSE and train MSE  i never said i wanted to generalize between public LB and CV . Other kernels are just overfitting on purpose as otherwise they would have done a train valid split to measure it i have reduced and simplified the LGDM model to fix overfit</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1599871,
      "author_name": "vladimircape",
      "author_url": "",
      "post_date": "11/29/2021 22:48:16",
      "content": "<p>90% of public notebooks made, to get additional votes as \"Notebooks Contributor\" and they don't have big benefits for users.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1581267": "Hi ,\n\nI have been into kaggle and machine learning for more than a year but i have never seen something like this thanks to the targets of the public lb present in train.csv . I noticed that most of the xgb and lgdm kernels were not doing something very important and common what say obvious people dont split between a validation and a train set like why ? I think as the targets of public lb are there in  ```train.csv``` they dont care about overfit and dont split the data and just overfit to the targets of public LB or overfitting to public lb in general . I think this is the reason though is foolish as they will just fail in private lb i want to listen to other peoples thoughts on this\n\nThank you",
    "1581426": "I agree with you. However,  FYI you haven't excluded  the overlapping public LB samples from your kernel (LB 0.638) as well, so the score doesn't reflect the true generalisation performance of the model.",
    "1582125": "Hi,\nit is always great to have some extra data from public lb or extra data in general i did not create that notebook for public lb extra data may improve overfit focus that i mean 'generalization' which means between Validation MSE and train MSE  i never said i wanted to generalize between public LB and CV . Other kernels are just overfitting on purpose as otherwise they would have done a train valid split to measure it i have reduced and simplified the LGDM model to fix overfit",
    "1599871": "90% of public notebooks made, to get additional votes as \"Notebooks Contributor\" and they don't have big benefits for users."
  },
  "source": "meta"
}