{
  "id": 671766,
  "title": "My first proper competition",
  "url": "/competitions/cafa-6-protein-function-prediction/discussion/671766",
  "author_name": "",
  "post_date": "2026-02-03T22:21:45.378671Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>This has been my first proper Kaggle competition and I just wanted to thank the organisers, I really enjoyed the competition!</p>\n<p>My background is in medicinal chemistry - I only got into writing code at the beginning of 2025, and then got a bit obsessed with learning ML. For this competition I had reasonably deep domain knowledge, but at the start I pretty much didn't even know where to begin with the ML aspect. I have learned so much, tried so many things that often didn't work , failed a lot, learned even more, and gradually improved and ended up with something that is performing pretty well!</p>\n<p>How do these things usually work after the deadline? I would really like to know things that other people tried. What were peoples thought processes etc? How did people deal with issues that came up? One of the things I haven't fully figured out yet is I made a range of diverse models and they give me a range of confidences but since the F1max is based on ranking I didn't know how to best calibrate the outputs. Some models might be inherently more confident so I was dubious about just normalising. I also found weighted ensembles to be better than XGboost (I tried loads of feature engineering - it always just seemed to learn that one of the models or features was dominant).</p>\n<p>Thank you again organisers :)</p>",
  "messages": [
    {
      "id": "3401519",
      "postDate": "02/03/2026 22:21:45",
      "content": "<p>This has been my first proper Kaggle competition and I just wanted to thank the organisers, I really enjoyed the competition!</p>\n<p>My background is in medicinal chemistry - I only got into writing code at the beginning of 2025, and then got a bit obsessed with learning ML. For this competition I had reasonably deep domain knowledge, but at the start I pretty much didn't even know where to begin with the ML aspect. I have learned so much, tried so many things that often didn't work , failed a lot, learned even more, and gradually improved and ended up with something that is performing pretty well!</p>\n<p>How do these things usually work after the deadline? I would really like to know things that other people tried. What were peoples thought processes etc? How did people deal with issues that came up? One of the things I haven't fully figured out yet is I made a range of diverse models and they give me a range of confidences but since the F1max is based on ranking I didn't know how to best calibrate the outputs. Some models might be inherently more confident so I was dubious about just normalising. I also found weighted ensembles to be better than XGboost (I tried loads of feature engineering - it always just seemed to learn that one of the models or features was dominant).</p>\n<p>Thank you again organisers :)</p>",
      "rawMarkdown": "This has been my first proper Kaggle competition and I just wanted to thank the organisers, I really enjoyed the competition!\n\nMy background is in medicinal chemistry - I only got into writing code at the beginning of 2025, and then got a bit obsessed with learning ML. For this competition I had reasonably deep domain knowledge, but at the start I pretty much didn't even know where to begin with the ML aspect. I have learned so much, tried so many things that often didn't work , failed a lot, learned even more, and gradually improved and ended up with something that is performing pretty well!\n\nHow do these things usually work after the deadline? I would really like to know things that other people tried. What were peoples thought processes etc? How did people deal with issues that came up? One of the things I haven't fully figured out yet is I made a range of diverse models and they give me a range of confidences but since the F1max is based on ranking I didn't know how to best calibrate the outputs. Some models might be inherently more confident so I was dubious about just normalising. I also found weighted ensembles to be better than XGboost (I tried loads of feature engineering - it always just seemed to learn that one of the models or features was dominant).\n\nThank you again organisers :)",
      "votes": null
    },
    {
      "id": "3401524",
      "postDate": "02/03/2026 22:45:11",
      "content": "<p>Glad to hear that you enjoyed the competition!</p>\n<p>After the submission deadline, we will wait for new GO terms to accumulate. For every batch of GO annotations released to the public (usually overlap with a UniProt release roughly every 8 weeks), we get new ground truth and evaluate all the submissions on this new ground truth, then you can see the new leaderboard score.</p>",
      "rawMarkdown": "Glad to hear that you enjoyed the competition!\n\nAfter the submission deadline, we will wait for new GO terms to accumulate. For every batch of GO annotations released to the public (usually overlap with a UniProt release roughly every 8 weeks), we get new ground truth and evaluate all the submissions on this new ground truth, then you can see the new leaderboard score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3401524,
      "author_name": "ahphan",
      "author_url": "",
      "post_date": "02/03/2026 22:45:11",
      "content": "<p>Glad to hear that you enjoyed the competition!</p>\n<p>After the submission deadline, we will wait for new GO terms to accumulate. For every batch of GO annotations released to the public (usually overlap with a UniProt release roughly every 8 weeks), we get new ground truth and evaluate all the submissions on this new ground truth, then you can see the new leaderboard score.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3401519": "This has been my first proper Kaggle competition and I just wanted to thank the organisers, I really enjoyed the competition!\n\nMy background is in medicinal chemistry - I only got into writing code at the beginning of 2025, and then got a bit obsessed with learning ML. For this competition I had reasonably deep domain knowledge, but at the start I pretty much didn't even know where to begin with the ML aspect. I have learned so much, tried so many things that often didn't work , failed a lot, learned even more, and gradually improved and ended up with something that is performing pretty well!\n\nHow do these things usually work after the deadline? I would really like to know things that other people tried. What were peoples thought processes etc? How did people deal with issues that came up? One of the things I haven't fully figured out yet is I made a range of diverse models and they give me a range of confidences but since the F1max is based on ranking I didn't know how to best calibrate the outputs. Some models might be inherently more confident so I was dubious about just normalising. I also found weighted ensembles to be better than XGboost (I tried loads of feature engineering - it always just seemed to learn that one of the models or features was dominant).\n\nThank you again organisers :)",
    "3401524": "Glad to hear that you enjoyed the competition!\n\nAfter the submission deadline, we will wait for new GO terms to accumulate. For every batch of GO annotations released to the public (usually overlap with a UniProt release roughly every 8 weeks), we get new ground truth and evaluate all the submissions on this new ground truth, then you can see the new leaderboard score."
  },
  "source": "meta"
}