{
  "id": 614121,
  "title": "importance of metadata e.g. `target_name` for `target` value",
  "url": "/competitions/csiro-biomass/discussion/614121",
  "author_name": "",
  "post_date": "2025-11-01T08:43:20.463133500Z",
  "votes": 18,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I was quite surprised that if you take the median for each <code>target_name</code>, you would obtain even better results than many baseline models which ignore this metadata…</p>\n<p><strong>simple mean</strong>: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-simple-stat-predictions\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-simple-stat-predictions</a> | LB: 0.23</p>\n<p><strong>baseline</strong>: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-end-to-end-lit-timm\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-end-to-end-lit-timm</a> | LB: 0.08</p>\n<h2>So thinking about a few directions:</h2>\n<ul>\n<li>Use a pretrained model with frozen weights and just add a regression head with metadata as another input, see: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-xgboost-pca-image-features\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-xgboost-pca-image-features</a> | LB: 0.50</li>\n<li>Build and train a model that would, as part of the process, try to fit all provided metadata, so hopefully better guide the training of the underlying model, see: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-multi-head-lit-timm\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-multi-head-lit-timm</a></li>\n</ul>",
  "messages": [
    {
      "id": "3309734",
      "postDate": "11/01/2025 08:43:20",
      "content": "<p>I was quite surprised that if you take the median for each <code>target_name</code>, you would obtain even better results than many baseline models which ignore this metadata…</p>\n<p><strong>simple mean</strong>: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-simple-stat-predictions\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-simple-stat-predictions</a> | LB: 0.23</p>\n<p><strong>baseline</strong>: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-end-to-end-lit-timm\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-end-to-end-lit-timm</a> | LB: 0.08</p>\n<h2>So thinking about a few directions:</h2>\n<ul>\n<li>Use a pretrained model with frozen weights and just add a regression head with metadata as another input, see: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-xgboost-pca-image-features\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-xgboost-pca-image-features</a> | LB: 0.50</li>\n<li>Build and train a model that would, as part of the process, try to fit all provided metadata, so hopefully better guide the training of the underlying model, see: <a href=\"https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-multi-head-lit-timm\" target=\"_blank\">https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-multi-head-lit-timm</a></li>\n</ul>",
      "rawMarkdown": "I was quite surprised that if you take the median for each `target_name`, you would obtain even better results than many baseline models which ignore this metadata...\n\n**simple mean**: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-simple-stat-predictions | LB: 0.23\n\n**baseline**: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-end-to-end-lit-timm | LB: 0.08\n\n## So thinking about a few directions:\n\n- Use a pretrained model with frozen weights and just add a regression head with metadata as another input, see: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-xgboost-pca-image-features | LB: 0.50\n- Build and train a model that would, as part of the process, try to fit all provided metadata, so hopefully better guide the training of the underlying model, see: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-multi-head-lit-timm",
      "votes": null
    },
    {
      "id": "3309832",
      "postDate": "11/01/2025 14:56:28",
      "content": "<p>I first thought about making a multimodal for incorporating the tabular data, but saw that the same data is not present in the test.</p>\n<p>The first suggested approach from what I understand it is, would hence not work.\nThis 2nd approach though is really clever. Making a model having to predict this additional information can surely make it capture some features it would have otherwise ignored. To make this happen, I believe the backbone has to be training, i.e. not frozen. 🧊 \nRectify my understanding.</p>\n<p><strong>Predicting these features can be treated as auxiliary tasks.</strong></p>",
      "rawMarkdown": "I first thought about making a multimodal for incorporating the tabular data, but saw that the same data is not present in the test.\n\nThe first suggested approach from what I understand it is, would hence not work.\nThis 2nd approach though is really clever. Making a model having to predict this additional information can surely make it capture some features it would have otherwise ignored. To make this happen, I believe the backbone has to be training, i.e. not frozen. 🧊 \nRectify my understanding.\n\n**Predicting these features can be treated as auxiliary tasks.**",
      "votes": null
    },
    {
      "id": "3309928",
      "postDate": "11/01/2025 18:10:00",
      "content": "<p>The reference paper appears to have used only the tabular data.  The models look like they are getting  results similar to the current leaderboard.   Based on the poor quality of the images it seems they never intended to use images early in the process of experiment design.  If they had, you would think they would have taken more care to keep things like the tripod shadows  out of the picture.  I plan to spend a decent amount of time figuring out the best approach to handle the artifacts in the images before I try to process the metadata.    </p>",
      "rawMarkdown": "The reference paper appears to have used only the tabular data.  The models look like they are getting  results similar to the current leaderboard.   Based on the poor quality of the images it seems they never intended to use images early in the process of experiment design.  If they had, you would think they would have taken more care to keep things like the tripod shadows  out of the picture.  I plan to spend a decent amount of time figuring out the best approach to handle the artifacts in the images before I try to process the metadata.",
      "votes": null
    },
    {
      "id": "3309996",
      "postDate": "11/01/2025 20:42:29",
      "content": "<p>Pretty much all now leading submissions has mix of image and metadata… </p>",
      "rawMarkdown": "Pretty much all now leading submissions has mix of image and metadata...",
      "votes": null
    },
    {
      "id": "3310600",
      "postDate": "11/03/2025 08:54:11",
      "content": "<p>So I have added image features with Dino V3 ConvNext to plain Metadata and jumped from 0.23 to 0.50</p>",
      "rawMarkdown": "So I have added image features with Dino V3 ConvNext to plain Metadata and jumped from 0.23 to 0.50",
      "votes": null
    },
    {
      "id": "3326689",
      "postDate": "11/15/2025 11:31:46",
      "content": "<p>I learned a lot.Good luck!</p>",
      "rawMarkdown": "I learned a lot.Good luck!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3309832,
      "author_name": "komilparmar",
      "author_url": "",
      "post_date": "11/01/2025 14:56:28",
      "content": "<p>I first thought about making a multimodal for incorporating the tabular data, but saw that the same data is not present in the test.</p>\n<p>The first suggested approach from what I understand it is, would hence not work.\nThis 2nd approach though is really clever. Making a model having to predict this additional information can surely make it capture some features it would have otherwise ignored. To make this happen, I believe the backbone has to be training, i.e. not frozen. 🧊 \nRectify my understanding.</p>\n<p><strong>Predicting these features can be treated as auxiliary tasks.</strong></p>",
      "votes": null,
      "replies": [
        {
          "id": 3309996,
          "author_name": "jirkaborovec",
          "author_url": "",
          "post_date": "11/01/2025 20:42:29",
          "content": "<p>Pretty much all now leading submissions has mix of image and metadata… </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3309928,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "11/01/2025 18:10:00",
      "content": "<p>The reference paper appears to have used only the tabular data.  The models look like they are getting  results similar to the current leaderboard.   Based on the poor quality of the images it seems they never intended to use images early in the process of experiment design.  If they had, you would think they would have taken more care to keep things like the tripod shadows  out of the picture.  I plan to spend a decent amount of time figuring out the best approach to handle the artifacts in the images before I try to process the metadata.    </p>",
      "votes": null,
      "replies": [
        {
          "id": 3326689,
          "author_name": "jakkma",
          "author_url": "",
          "post_date": "11/15/2025 11:31:46",
          "content": "<p>I learned a lot.Good luck!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3310600,
      "author_name": "jirkaborovec",
      "author_url": "",
      "post_date": "11/03/2025 08:54:11",
      "content": "<p>So I have added image features with Dino V3 ConvNext to plain Metadata and jumped from 0.23 to 0.50</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3309734": "I was quite surprised that if you take the median for each `target_name`, you would obtain even better results than many baseline models which ignore this metadata...\n\n**simple mean**: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-simple-stat-predictions | LB: 0.23\n\n**baseline**: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-end-to-end-lit-timm | LB: 0.08\n\n## So thinking about a few directions:\n\n- Use a pretrained model with frozen weights and just add a regression head with metadata as another input, see: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-xgboost-pca-image-features | LB: 0.50\n- Build and train a model that would, as part of the process, try to fit all provided metadata, so hopefully better guide the training of the underlying model, see: https://www.kaggle.com/code/jirkaborovec/csiro-img2bio-eda-multi-head-lit-timm",
    "3309832": "I first thought about making a multimodal for incorporating the tabular data, but saw that the same data is not present in the test.\n\nThe first suggested approach from what I understand it is, would hence not work.\nThis 2nd approach though is really clever. Making a model having to predict this additional information can surely make it capture some features it would have otherwise ignored. To make this happen, I believe the backbone has to be training, i.e. not frozen. 🧊 \nRectify my understanding.\n\n**Predicting these features can be treated as auxiliary tasks.**",
    "3309928": "The reference paper appears to have used only the tabular data.  The models look like they are getting  results similar to the current leaderboard.   Based on the poor quality of the images it seems they never intended to use images early in the process of experiment design.  If they had, you would think they would have taken more care to keep things like the tripod shadows  out of the picture.  I plan to spend a decent amount of time figuring out the best approach to handle the artifacts in the images before I try to process the metadata.",
    "3309996": "Pretty much all now leading submissions has mix of image and metadata...",
    "3310600": "So I have added image features with Dino V3 ConvNext to plain Metadata and jumped from 0.23 to 0.50",
    "3326689": "I learned a lot.Good luck!"
  },
  "source": "meta"
}