{
  "id": 614190,
  "title": "Metric ambiguity",
  "url": "/competitions/csiro-biomass/discussion/614190",
  "author_name": "",
  "post_date": "2025-11-01T19:21:19.505539900Z",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi. Competition page doesn't say anything about log transformation but in this related paper <a href=\"https://arxiv.org/html/2510.22916v1\" target=\"_blank\">link</a>, it is mentioned. Which is right?</p>",
  "messages": [
    {
      "id": "3309965",
      "postDate": "11/01/2025 19:21:19",
      "content": "<p>Hi. Competition page doesn't say anything about log transformation but in this related paper <a href=\"https://arxiv.org/html/2510.22916v1\" target=\"_blank\">link</a>, it is mentioned. Which is right?</p>",
      "rawMarkdown": "Hi. Competition page doesn't say anything about log transformation but in this related paper [link](https://arxiv.org/html/2510.22916v1), it is mentioned. Which is right?",
      "votes": null
    },
    {
      "id": "3309988",
      "postDate": "11/01/2025 20:10:28",
      "content": "<p>The paper is not an official reference for this competition. We don’t use any log transformation here.</p>",
      "rawMarkdown": "The paper is not an official reference for this competition. We don’t use any log transformation here.",
      "votes": null
    },
    {
      "id": "3310032",
      "postDate": "11/01/2025 23:08:23",
      "content": "<p>Good question - thanks for asking.   It might make sense to use the log transformation for the target values when building your model - just need to remember to convert predictions back.  The plot of my most recent models predictions on the full train set does have the kind of shape that is often helped by building the model with log transformed data.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1062256%2F3350ed9db6caad2e77b2c140a1c341fc%2FR015_Model_Fit.png?generation=1762038432708831&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "Good question - thanks for asking.   It might make sense to use the log transformation for the target values when building your model - just need to remember to convert predictions back.  The plot of my most recent models predictions on the full train set does have the kind of shape that is often helped by building the model with log transformed data.  \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1062256%2F3350ed9db6caad2e77b2c140a1c341fc%2FR015_Model_Fit.png?generation=1762038432708831&alt=media)",
      "votes": null
    },
    {
      "id": "3310070",
      "postDate": "11/02/2025 02:44:10",
      "content": "<p>It can also be seen that the target distribution has a long tail curve, so log transform should be helpful.</p>",
      "rawMarkdown": "It can also be seen that the target distribution has a long tail curve, so log transform should be helpful.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3309988,
      "author_name": "martynaplomecka",
      "author_url": "",
      "post_date": "11/01/2025 20:10:28",
      "content": "<p>The paper is not an official reference for this competition. We don’t use any log transformation here.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3310032,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "11/01/2025 23:08:23",
      "content": "<p>Good question - thanks for asking.   It might make sense to use the log transformation for the target values when building your model - just need to remember to convert predictions back.  The plot of my most recent models predictions on the full train set does have the kind of shape that is often helped by building the model with log transformed data.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1062256%2F3350ed9db6caad2e77b2c140a1c341fc%2FR015_Model_Fit.png?generation=1762038432708831&amp;alt=media\" alt=\"\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 3310070,
          "author_name": "paritoshtripathi5",
          "author_url": "",
          "post_date": "11/02/2025 02:44:10",
          "content": "<p>It can also be seen that the target distribution has a long tail curve, so log transform should be helpful.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3309965": "Hi. Competition page doesn't say anything about log transformation but in this related paper [link](https://arxiv.org/html/2510.22916v1), it is mentioned. Which is right?",
    "3309988": "The paper is not an official reference for this competition. We don’t use any log transformation here.",
    "3310032": "Good question - thanks for asking.   It might make sense to use the log transformation for the target values when building your model - just need to remember to convert predictions back.  The plot of my most recent models predictions on the full train set does have the kind of shape that is often helped by building the model with log transformed data.  \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1062256%2F3350ed9db6caad2e77b2c140a1c341fc%2FR015_Model_Fit.png?generation=1762038432708831&alt=media)",
    "3310070": "It can also be seen that the target distribution has a long tail curve, so log transform should be helpful."
  },
  "source": "meta"
}