{
  "id": 580206,
  "title": "Ensemble of solutions",
  "url": "/competitions/stanford-rna-3d-folding/discussion/580206",
  "author_name": "",
  "post_date": "2025-05-23T06:17:51.247460400Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>When i use single solution i got score of 0.35 but when i ensemble two solution with weight average the score fall to 0.31.So, which technique of ensemble performs better here? or should i avoid ensemble?</p>",
  "messages": [
    {
      "id": "3207692",
      "postDate": "05/23/2025 06:17:51",
      "content": "<p>When i use single solution i got score of 0.35 but when i ensemble two solution with weight average the score fall to 0.31.So, which technique of ensemble performs better here? or should i avoid ensemble?</p>",
      "rawMarkdown": "When i use single solution i got score of 0.35 but when i ensemble two solution with weight average the score fall to 0.31.So, which technique of ensemble performs better here? or should i avoid ensemble?",
      "votes": null
    },
    {
      "id": "3207717",
      "postDate": "05/23/2025 06:54:07",
      "content": "<p>I have met the same situation before, you can try to merge the tr model based on protenix, more hint in shared code.</p>",
      "rawMarkdown": "I have met the same situation before, you can try to merge the tr model based on protenix, more hint in shared code.",
      "votes": null
    },
    {
      "id": "3207936",
      "postDate": "05/23/2025 12:50:37",
      "content": "<p>The predictions from a model can't be ensembled straight away like, for example, taking the mean of each row. This is because sequences must be aligned beforehand. For instance, there might be two very structurally-similar sequences which are rotated or translated, so if you take the average it will be trash. It might be helpful if you first align all sequences against a common axis (mathematically, a rigid transformation rotation+translation) and then perform the ensemble. <a href=\"https://www.kaggle.com/swandipsingha\" target=\"_blank\">@swandipsingha</a> </p>",
      "rawMarkdown": "The predictions from a model can't be ensembled straight away like, for example, taking the mean of each row. This is because sequences must be aligned beforehand. For instance, there might be two very structurally-similar sequences which are rotated or translated, so if you take the average it will be trash. It might be helpful if you first align all sequences against a common axis (mathematically, a rigid transformation rotation+translation) and then perform the ensemble. @swandipsingha",
      "votes": null
    },
    {
      "id": "3207948",
      "postDate": "05/23/2025 13:02:38",
      "content": "<p><a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a> Thanks man</p>",
      "rawMarkdown": "alejopaullier Thanks man",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3207717,
      "author_name": "",
      "author_url": "",
      "post_date": "05/23/2025 06:54:07",
      "content": "<p>I have met the same situation before, you can try to merge the tr model based on protenix, more hint in shared code.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3207936,
      "author_name": "alejopaullier",
      "author_url": "",
      "post_date": "05/23/2025 12:50:37",
      "content": "<p>The predictions from a model can't be ensembled straight away like, for example, taking the mean of each row. This is because sequences must be aligned beforehand. For instance, there might be two very structurally-similar sequences which are rotated or translated, so if you take the average it will be trash. It might be helpful if you first align all sequences against a common axis (mathematically, a rigid transformation rotation+translation) and then perform the ensemble. <a href=\"https://www.kaggle.com/swandipsingha\" target=\"_blank\">@swandipsingha</a> </p>",
      "votes": null,
      "replies": [
        {
          "id": 3207948,
          "author_name": "swandipsingha",
          "author_url": "",
          "post_date": "05/23/2025 13:02:38",
          "content": "<p><a href=\"https://www.kaggle.com/alejopaullier\" target=\"_blank\">@alejopaullier</a> Thanks man</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3207692": "When i use single solution i got score of 0.35 but when i ensemble two solution with weight average the score fall to 0.31.So, which technique of ensemble performs better here? or should i avoid ensemble?",
    "3207717": "I have met the same situation before, you can try to merge the tr model based on protenix, more hint in shared code.",
    "3207936": "The predictions from a model can't be ensembled straight away like, for example, taking the mean of each row. This is because sequences must be aligned beforehand. For instance, there might be two very structurally-similar sequences which are rotated or translated, so if you take the average it will be trash. It might be helpful if you first align all sequences against a common axis (mathematically, a rigid transformation rotation+translation) and then perform the ensemble. @swandipsingha",
    "3207948": "alejopaullier Thanks man"
  },
  "source": "meta"
}