{
  "id": 410018,
  "title": "Annotations on log-scaled axes are wrong",
  "url": "/competitions/benetech-making-graphs-accessible/discussion/410018",
  "author_name": "",
  "post_date": "2023-05-13T14:51:52.857962Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Take the example of <code>id=b2febc0f5a82</code>, which is extracted i.e. could be in the test set.</p>\n<p><a href=\"https://ibb.co/g97pt59\"><img src=\"https://i.ibb.co/wgLXp9g/logscale.png\" alt=\"logscale\"></a></p>\n<p>We have clusters of 2 points at <code>x=2.5</code>, <code>x=5</code>, <code>x=10</code>, <code>x=20</code> and <code>x=40</code>.</p>\n<p>However, the ground truth is :</p>\n<pre><code>[{'x': 5, 'y': 0.99...}, {'x': 5, 'y': 0.94...}, {'x': 7, 'y': 0.97...}, {'x': 7, 'y': 0.96...}, {'x': 10, 'y': 1.01...}, {'x': 10, 'y': 0.98...}, {'x': 37, 'y': 1.01...}, {'x': 37, 'y': 1.0...}, {'x': 63, 'y': 1.11...}, {'x': 64, 'y': 1.10...}, {'x': 64, 'y': 1.06...}]\n</code></pre>\n<p>Which is close to what you would get by linearly interpolating between ticks, and is wrong.<br>\nI have not checked for all the images, but this was the case for the 3 scatterplots I found. I'm deducing that there is an issue in the labeling pipeline, and therefore all log values could be wrong. </p>\n<p>EDIT : Also note that  models would score 0 on this chart, because there are two points overlapping that I did not even spot =)</p>\n<p>Update: Already discussed, see comments </p>",
  "messages": [
    {
      "id": "2257645",
      "postDate": "05/13/2023 14:51:52",
      "content": "<p>Take the example of <code>id=b2febc0f5a82</code>, which is extracted i.e. could be in the test set.</p>\n<p><a href=\"https://ibb.co/g97pt59\"><img src=\"https://i.ibb.co/wgLXp9g/logscale.png\" alt=\"logscale\"></a></p>\n<p>We have clusters of 2 points at <code>x=2.5</code>, <code>x=5</code>, <code>x=10</code>, <code>x=20</code> and <code>x=40</code>.</p>\n<p>However, the ground truth is :</p>\n<pre><code>[{'x': 5, 'y': 0.99...}, {'x': 5, 'y': 0.94...}, {'x': 7, 'y': 0.97...}, {'x': 7, 'y': 0.96...}, {'x': 10, 'y': 1.01...}, {'x': 10, 'y': 0.98...}, {'x': 37, 'y': 1.01...}, {'x': 37, 'y': 1.0...}, {'x': 63, 'y': 1.11...}, {'x': 64, 'y': 1.10...}, {'x': 64, 'y': 1.06...}]\n</code></pre>\n<p>Which is close to what you would get by linearly interpolating between ticks, and is wrong.<br>\nI have not checked for all the images, but this was the case for the 3 scatterplots I found. I'm deducing that there is an issue in the labeling pipeline, and therefore all log values could be wrong. </p>\n<p>EDIT : Also note that  models would score 0 on this chart, because there are two points overlapping that I did not even spot =)</p>\n<p>Update: Already discussed, see comments </p>",
      "rawMarkdown": "Take the example of `id=b2febc0f5a82`, which is extracted i.e. could be in the test set.\n\n<a href=\"https://ibb.co/g97pt59\"><img src=\"https://i.ibb.co/wgLXp9g/logscale.png\" alt=\"logscale\" border=\"0\"></a>\n\nWe have clusters of 2 points at `x=2.5`, `x=5`, `x=10`, `x=20` and `x=40`.\n\nHowever, the ground truth is :\n\n```\n[{'x': 5, 'y': 0.99...}, {'x': 5, 'y': 0.94...}, {'x': 7, 'y': 0.97...}, {'x': 7, 'y': 0.96...}, {'x': 10, 'y': 1.01...}, {'x': 10, 'y': 0.98...}, {'x': 37, 'y': 1.01...}, {'x': 37, 'y': 1.0...}, {'x': 63, 'y': 1.11...}, {'x': 64, 'y': 1.10...}, {'x': 64, 'y': 1.06...}]\n```\n\nWhich is close to what you would get by linearly interpolating between ticks, and is wrong.\nI have not checked for all the images, but this was the case for the 3 scatterplots I found. I'm deducing that there is an issue in the labeling pipeline, and therefore all log values could be wrong. \n\nEDIT : Also note that  models would score 0 on this chart, because there are two points overlapping that I did not even spot =)\n\nUpdate: Already discussed, see comments",
      "votes": null
    },
    {
      "id": "2257871",
      "postDate": "05/13/2023 17:52:26",
      "content": "<p>There are no log-scale charts in the test. This point was previously <a href=\"https://www.kaggle.com/competitions/benetech-making-graphs-accessible/discussion/398601\" target=\"_blank\">discussed</a>.</p>",
      "rawMarkdown": "There are no log-scale charts in the test. This point was previously [discussed](https://www.kaggle.com/competitions/benetech-making-graphs-accessible/discussion/398601).",
      "votes": null
    },
    {
      "id": "2258484",
      "postDate": "05/14/2023 08:47:42",
      "content": "<p>Thanks, I missed that !</p>",
      "rawMarkdown": "Thanks, I missed that !",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2257871,
      "author_name": "awqatak",
      "author_url": "",
      "post_date": "05/13/2023 17:52:26",
      "content": "<p>There are no log-scale charts in the test. This point was previously <a href=\"https://www.kaggle.com/competitions/benetech-making-graphs-accessible/discussion/398601\" target=\"_blank\">discussed</a>.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2258484,
          "author_name": "theoviel",
          "author_url": "",
          "post_date": "05/14/2023 08:47:42",
          "content": "<p>Thanks, I missed that !</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2257645": "Take the example of `id=b2febc0f5a82`, which is extracted i.e. could be in the test set.\n\n<a href=\"https://ibb.co/g97pt59\"><img src=\"https://i.ibb.co/wgLXp9g/logscale.png\" alt=\"logscale\" border=\"0\"></a>\n\nWe have clusters of 2 points at `x=2.5`, `x=5`, `x=10`, `x=20` and `x=40`.\n\nHowever, the ground truth is :\n\n```\n[{'x': 5, 'y': 0.99...}, {'x': 5, 'y': 0.94...}, {'x': 7, 'y': 0.97...}, {'x': 7, 'y': 0.96...}, {'x': 10, 'y': 1.01...}, {'x': 10, 'y': 0.98...}, {'x': 37, 'y': 1.01...}, {'x': 37, 'y': 1.0...}, {'x': 63, 'y': 1.11...}, {'x': 64, 'y': 1.10...}, {'x': 64, 'y': 1.06...}]\n```\n\nWhich is close to what you would get by linearly interpolating between ticks, and is wrong.\nI have not checked for all the images, but this was the case for the 3 scatterplots I found. I'm deducing that there is an issue in the labeling pipeline, and therefore all log values could be wrong. \n\nEDIT : Also note that  models would score 0 on this chart, because there are two points overlapping that I did not even spot =)\n\nUpdate: Already discussed, see comments",
    "2257871": "There are no log-scale charts in the test. This point was previously [discussed](https://www.kaggle.com/competitions/benetech-making-graphs-accessible/discussion/398601).",
    "2258484": "Thanks, I missed that !"
  },
  "source": "meta"
}