{
  "id": 662651,
  "title": "K-fold or single CV?",
  "url": "/competitions/csiro-biomass/discussion/662651",
  "author_name": "",
  "post_date": "2025-12-14T03:59:46.969038400Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi! Thanks for Kaggle - the challenges are really motivating for me. </p>\n<p>Could you advise me on how to use training data with k-fold cross-validation, or with a single train/validation split? With a small dataset, so I’m concerned that k-fold might lead to overfitting. </p>\n<p>Thanks for any helpful advice.</p>",
  "messages": [
    {
      "id": "3376278",
      "postDate": "12/14/2025 03:59:46",
      "content": "<p>Hi! Thanks for Kaggle - the challenges are really motivating for me. </p>\n<p>Could you advise me on how to use training data with k-fold cross-validation, or with a single train/validation split? With a small dataset, so I’m concerned that k-fold might lead to overfitting. </p>\n<p>Thanks for any helpful advice.</p>",
      "rawMarkdown": "Hi! Thanks for Kaggle - the challenges are really motivating for me. \n\nCould you advise me on how to use training data with k-fold cross-validation, or with a single train/validation split? With a small dataset, so I’m concerned that k-fold might lead to overfitting. \n\nThanks for any helpful advice.",
      "votes": null
    },
    {
      "id": "3376285",
      "postDate": "12/14/2025 04:16:59",
      "content": "<p>None actually. Try each and you will know why - if you do it so simply than there will be data leakage in various forms like site, temporal etc which will make your models get high CV scores easily but you won't get comparable LB score. </p>\n<p>I recommend you thoroughly go through the current highest rated discussion. \n<a href=\"https://www.kaggle.com/competitions/csiro-biomass/discussion/615401\" target=\"_blank\">https://www.kaggle.com/competitions/csiro-biomass/discussion/615401</a></p>",
      "rawMarkdown": "None actually. Try each and you will know why - if you do it so simply than there will be data leakage in various forms like site, temporal etc which will make your models get high CV scores easily but you won't get comparable LB score. \n\nI recommend you thoroughly go through the current highest rated discussion. \nhttps://www.kaggle.com/competitions/csiro-biomass/discussion/615401",
      "votes": null
    },
    {
      "id": "3376290",
      "postDate": "12/14/2025 04:25:17",
      "content": "<p>Thank you. I’ve been struggling with how to implement K-fold, so I’ll consider your approach as a reference</p>",
      "rawMarkdown": "Thank you. I’ve been struggling with how to implement K-fold, so I’ll consider your approach as a reference",
      "votes": null
    },
    {
      "id": "3376304",
      "postDate": "12/14/2025 05:39:57",
      "content": "<p>With a small data set you almost have to do K folds otherwise you miss a sizable part of the information.  Besides being small there's lots of evidence that it's got lots of measurement errors also.</p>",
      "rawMarkdown": "With a small data set you almost have to do K folds otherwise you miss a sizable part of the information.  Besides being small there's lots of evidence that it's got lots of measurement errors also.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3376285,
      "author_name": "komilparmar",
      "author_url": "",
      "post_date": "12/14/2025 04:16:59",
      "content": "<p>None actually. Try each and you will know why - if you do it so simply than there will be data leakage in various forms like site, temporal etc which will make your models get high CV scores easily but you won't get comparable LB score. </p>\n<p>I recommend you thoroughly go through the current highest rated discussion. \n<a href=\"https://www.kaggle.com/competitions/csiro-biomass/discussion/615401\" target=\"_blank\">https://www.kaggle.com/competitions/csiro-biomass/discussion/615401</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 3376290,
          "author_name": "taiki2",
          "author_url": "",
          "post_date": "12/14/2025 04:25:17",
          "content": "<p>Thank you. I’ve been struggling with how to implement K-fold, so I’ll consider your approach as a reference</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3376304,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "12/14/2025 05:39:57",
      "content": "<p>With a small data set you almost have to do K folds otherwise you miss a sizable part of the information.  Besides being small there's lots of evidence that it's got lots of measurement errors also.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3376278": "Hi! Thanks for Kaggle - the challenges are really motivating for me. \n\nCould you advise me on how to use training data with k-fold cross-validation, or with a single train/validation split? With a small dataset, so I’m concerned that k-fold might lead to overfitting. \n\nThanks for any helpful advice.",
    "3376285": "None actually. Try each and you will know why - if you do it so simply than there will be data leakage in various forms like site, temporal etc which will make your models get high CV scores easily but you won't get comparable LB score. \n\nI recommend you thoroughly go through the current highest rated discussion. \nhttps://www.kaggle.com/competitions/csiro-biomass/discussion/615401",
    "3376290": "Thank you. I’ve been struggling with how to implement K-fold, so I’ll consider your approach as a reference",
    "3376304": "With a small data set you almost have to do K folds otherwise you miss a sizable part of the information.  Besides being small there's lots of evidence that it's got lots of measurement errors also."
  },
  "source": "meta"
}