{
  "id": 570179,
  "title": "How do we ensemble models to outscore predictions generated by single model?",
  "url": "/competitions/stanford-rna-3d-folding/discussion/570179",
  "author_name": "",
  "post_date": "2025-03-26T10:36:23.030710Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Instead of putting each model's prediction separately into submission file, I'm thinking of USalign two predicted structure and then<br>\n・calculate average point for each corresponding atoms<br>\n・extract local feature by GCN→Cross-Attention→generate mixed structure by VAE(work on loss function to maintain some local features)</p>\n<p>Is there anyone already trying these things? Any thoughts on other ensembling methods?<br>\nDoes mixing prediction like this improve score?</p>",
  "messages": [
    {
      "id": "3160072",
      "postDate": "03/26/2025 10:36:23",
      "content": "<p>Instead of putting each model's prediction separately into submission file, I'm thinking of USalign two predicted structure and then<br>\n・calculate average point for each corresponding atoms<br>\n・extract local feature by GCN→Cross-Attention→generate mixed structure by VAE(work on loss function to maintain some local features)</p>\n<p>Is there anyone already trying these things? Any thoughts on other ensembling methods?<br>\nDoes mixing prediction like this improve score?</p>",
      "rawMarkdown": "Instead of putting each model's prediction separately into submission file, I'm thinking of USalign two predicted structure and then\n・calculate average point for each corresponding atoms\n・extract local feature by GCN→Cross-Attention→generate mixed structure by VAE(work on loss function to maintain some local features)\n\nIs there anyone already trying these things? Any thoughts on other ensembling methods?\nDoes mixing prediction like this improve score?",
      "votes": null
    },
    {
      "id": "3160172",
      "postDate": "03/26/2025 13:00:28",
      "content": "<p>For each sequence, you can submit 5 structures. If you have more than 5 results, you can use some clustering methods or calculate the energy to select the best 5</p>",
      "rawMarkdown": "For each sequence, you can submit 5 structures. If you have more than 5 results, you can use some clustering methods or calculate the energy to select the best 5",
      "votes": null
    },
    {
      "id": "3160180",
      "postDate": "03/26/2025 13:12:06",
      "content": "<p>Thanks for comment!<br>\n Is there any way submitting mixed(=like calculating average points/use gcns etc to generate “mixed structure) prediction can outscore prediction generated by a single model? Sorry for my bad English, and simply finding the best 5 as your advice might be the best strategy, but ensemble is my interest, so looking forward to your reply 🙇‍♀️</p>",
      "rawMarkdown": "Thanks for comment!\n Is there any way submitting mixed(=like calculating average points/use gcns etc to generate “mixed structure) prediction can outscore prediction generated by a single model? Sorry for my bad English, and simply finding the best 5 as your advice might be the best strategy, but ensemble is my interest, so looking forward to your reply 🙇‍♀️",
      "votes": null
    },
    {
      "id": "3160219",
      "postDate": "03/26/2025 14:11:09",
      "content": "<p>I haven't tried this so I am not sure. You may need to align these structures first and then calculate the average xyz? Be aware this may lead to substantial errors. </p>",
      "rawMarkdown": "I haven't tried this so I am not sure. You may need to align these structures first and then calculate the average xyz? Be aware this may lead to substantial errors.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3160172,
      "author_name": "biancochiu",
      "author_url": "",
      "post_date": "03/26/2025 13:00:28",
      "content": "<p>For each sequence, you can submit 5 structures. If you have more than 5 results, you can use some clustering methods or calculate the energy to select the best 5</p>",
      "votes": null,
      "replies": [
        {
          "id": 3160180,
          "author_name": "nanacat0520",
          "author_url": "",
          "post_date": "03/26/2025 13:12:06",
          "content": "<p>Thanks for comment!<br>\n Is there any way submitting mixed(=like calculating average points/use gcns etc to generate “mixed structure) prediction can outscore prediction generated by a single model? Sorry for my bad English, and simply finding the best 5 as your advice might be the best strategy, but ensemble is my interest, so looking forward to your reply 🙇‍♀️</p>",
          "votes": null,
          "replies": [
            {
              "id": 3160219,
              "author_name": "biancochiu",
              "author_url": "",
              "post_date": "03/26/2025 14:11:09",
              "content": "<p>I haven't tried this so I am not sure. You may need to align these structures first and then calculate the average xyz? Be aware this may lead to substantial errors. </p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3160072": "Instead of putting each model's prediction separately into submission file, I'm thinking of USalign two predicted structure and then\n・calculate average point for each corresponding atoms\n・extract local feature by GCN→Cross-Attention→generate mixed structure by VAE(work on loss function to maintain some local features)\n\nIs there anyone already trying these things? Any thoughts on other ensembling methods?\nDoes mixing prediction like this improve score?",
    "3160172": "For each sequence, you can submit 5 structures. If you have more than 5 results, you can use some clustering methods or calculate the energy to select the best 5",
    "3160180": "Thanks for comment!\n Is there any way submitting mixed(=like calculating average points/use gcns etc to generate “mixed structure) prediction can outscore prediction generated by a single model? Sorry for my bad English, and simply finding the best 5 as your advice might be the best strategy, but ensemble is my interest, so looking forward to your reply 🙇‍♀️",
    "3160219": "I haven't tried this so I am not sure. You may need to align these structures first and then calculate the average xyz? Be aware this may lead to substantial errors."
  },
  "source": "meta"
}