{
  "id": 572782,
  "title": "Training and test data is faulty!",
  "url": "/competitions/el-hackathon-2025/discussion/572782",
  "author_name": "",
  "post_date": "2025-04-11T09:50:18.839641100Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>It seems pretty clear that the spot positions are incorrect on all slides. Nevertheless, running xgboost on some basic features can get you a score of around 0.4-0.5. I'm guessing xgboost simply learns the average distribution of the abundances, and since there is a lot of bias you get this score of 0.4 for free. </p>\n<p>After manually correcting the spot positions, my cv is now 0.5-0.6 but test score is very low, like 0.24. Does anyone know why this is the case? </p>\n<p>I would strongly urge the organizers to fix the spot coordinates, as there is no reliable way of fixing the test spot coordinates unless you do it by hand and submit a bunch of times.</p>",
  "messages": [
    {
      "id": "3176386",
      "postDate": "04/11/2025 09:50:18",
      "content": "<p>It seems pretty clear that the spot positions are incorrect on all slides. Nevertheless, running xgboost on some basic features can get you a score of around 0.4-0.5. I'm guessing xgboost simply learns the average distribution of the abundances, and since there is a lot of bias you get this score of 0.4 for free. </p>\n<p>After manually correcting the spot positions, my cv is now 0.5-0.6 but test score is very low, like 0.24. Does anyone know why this is the case? </p>\n<p>I would strongly urge the organizers to fix the spot coordinates, as there is no reliable way of fixing the test spot coordinates unless you do it by hand and submit a bunch of times.</p>",
      "rawMarkdown": "It seems pretty clear that the spot positions are incorrect on all slides. Nevertheless, running xgboost on some basic features can get you a score of around 0.4-0.5. I'm guessing xgboost simply learns the average distribution of the abundances, and since there is a lot of bias you get this score of 0.4 for free. \n\nAfter manually correcting the spot positions, my cv is now 0.5-0.6 but test score is very low, like 0.24. Does anyone know why this is the case? \n\nI would strongly urge the organizers to fix the spot coordinates, as there is no reliable way of fixing the test spot coordinates unless you do it by hand and submit a bunch of times.",
      "votes": null
    },
    {
      "id": "3177316",
      "postDate": "04/12/2025 15:02:07",
      "content": "<p>The discussion about wrong spot positions: <a href=\"https://www.kaggle.com/competitions/el-hackathon-2025/discussion/569132\" target=\"_blank\">https://www.kaggle.com/competitions/el-hackathon-2025/discussion/569132</a><br>\nMaybe it'll help you</p>\n<p>Your CV is high but your test score is very low =&gt; Maybe overfiting problems or data shift problems between train and test<br>\nIf data shift =&gt; Try a new train-val split approach </p>",
      "rawMarkdown": "The discussion about wrong spot positions: https://www.kaggle.com/competitions/el-hackathon-2025/discussion/569132\nMaybe it'll help you\n\nYour CV is high but your test score is very low => Maybe overfiting problems or data shift problems between train and test\nIf data shift => Try a new train-val split approach",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3177316,
      "author_name": "hnglnh",
      "author_url": "",
      "post_date": "04/12/2025 15:02:07",
      "content": "<p>The discussion about wrong spot positions: <a href=\"https://www.kaggle.com/competitions/el-hackathon-2025/discussion/569132\" target=\"_blank\">https://www.kaggle.com/competitions/el-hackathon-2025/discussion/569132</a><br>\nMaybe it'll help you</p>\n<p>Your CV is high but your test score is very low =&gt; Maybe overfiting problems or data shift problems between train and test<br>\nIf data shift =&gt; Try a new train-val split approach </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3176386": "It seems pretty clear that the spot positions are incorrect on all slides. Nevertheless, running xgboost on some basic features can get you a score of around 0.4-0.5. I'm guessing xgboost simply learns the average distribution of the abundances, and since there is a lot of bias you get this score of 0.4 for free. \n\nAfter manually correcting the spot positions, my cv is now 0.5-0.6 but test score is very low, like 0.24. Does anyone know why this is the case? \n\nI would strongly urge the organizers to fix the spot coordinates, as there is no reliable way of fixing the test spot coordinates unless you do it by hand and submit a bunch of times.",
    "3177316": "The discussion about wrong spot positions: https://www.kaggle.com/competitions/el-hackathon-2025/discussion/569132\nMaybe it'll help you\n\nYour CV is high but your test score is very low => Maybe overfiting problems or data shift problems between train and test\nIf data shift => Try a new train-val split approach"
  },
  "source": "meta"
}