{
  "id": 400001,
  "title": "How many models to ensemble",
  "url": "/competitions/icecube-neutrinos-in-deep-ice/discussion/400001",
  "author_name": "",
  "post_date": "2023-04-06T12:21:06.756298900Z",
  "votes": 3,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Did anyone try to check the limits of ensembling for this dataset? How many models do you ensemble till you see diminishing returns?</p>",
  "messages": [
    {
      "id": "2211987",
      "postDate": "04/06/2023 12:21:06",
      "content": "<p>Did anyone try to check the limits of ensembling for this dataset? How many models do you ensemble till you see diminishing returns?</p>",
      "rawMarkdown": "Did anyone try to check the limits of ensembling for this dataset? How many models do you ensemble till you see diminishing returns?",
      "votes": null
    },
    {
      "id": "2213658",
      "postDate": "04/07/2023 18:37:49",
      "content": "<p>I find it hard to ensemble to models cause there are always errors that notebook threw exceptions when submitted to LB</p>",
      "rawMarkdown": "I find it hard to ensemble to models cause there are always errors that notebook threw exceptions when submitted to LB",
      "votes": null
    },
    {
      "id": "2214456",
      "postDate": "04/08/2023 13:44:31",
      "content": "<p>Why not test it on one training batch first to ensure no errors happen? </p>",
      "rawMarkdown": "Why not test it on one training batch first to ensure no errors happen?",
      "votes": null
    },
    {
      "id": "2217179",
      "postDate": "04/10/2023 17:03:55",
      "content": "<p>What models in particular have you ensembled?</p>",
      "rawMarkdown": "What models in particular have you ensembled?",
      "votes": null
    },
    {
      "id": "2217219",
      "postDate": "04/10/2023 17:34:05",
      "content": "<p>Three model ensembles seem good to me, but more do give slightly more improvement. And models trained with different losses seem to be good at different types of events.</p>",
      "rawMarkdown": "Three model ensembles seem good to me, but more do give slightly more improvement. And models trained with different losses seem to be good at different types of events.",
      "votes": null
    },
    {
      "id": "2221276",
      "postDate": "04/14/2023 05:39:07",
      "content": "<p>I think the answer depends on several factors, such as the diversity and quality of the models, the correlation between their predictions, and the weighting scheme for the ensemble. </p>",
      "rawMarkdown": "I think the answer depends on several factors, such as the diversity and quality of the models, the correlation between their predictions, and the weighting scheme for the ensemble.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2213658,
      "author_name": "xstargate",
      "author_url": "",
      "post_date": "04/07/2023 18:37:49",
      "content": "<p>I find it hard to ensemble to models cause there are always errors that notebook threw exceptions when submitted to LB</p>",
      "votes": null,
      "replies": [
        {
          "id": 2214456,
          "author_name": "dipamc77",
          "author_url": "",
          "post_date": "04/08/2023 13:44:31",
          "content": "<p>Why not test it on one training batch first to ensure no errors happen? </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2217179,
      "author_name": "domdejonge",
      "author_url": "",
      "post_date": "04/10/2023 17:03:55",
      "content": "<p>What models in particular have you ensembled?</p>",
      "votes": null,
      "replies": [
        {
          "id": 2217219,
          "author_name": "dipamc77",
          "author_url": "",
          "post_date": "04/10/2023 17:34:05",
          "content": "<p>Three model ensembles seem good to me, but more do give slightly more improvement. And models trained with different losses seem to be good at different types of events.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2221276,
      "author_name": "yus002",
      "author_url": "",
      "post_date": "04/14/2023 05:39:07",
      "content": "<p>I think the answer depends on several factors, such as the diversity and quality of the models, the correlation between their predictions, and the weighting scheme for the ensemble. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2211987": "Did anyone try to check the limits of ensembling for this dataset? How many models do you ensemble till you see diminishing returns?",
    "2213658": "I find it hard to ensemble to models cause there are always errors that notebook threw exceptions when submitted to LB",
    "2214456": "Why not test it on one training batch first to ensure no errors happen?",
    "2217179": "What models in particular have you ensembled?",
    "2217219": "Three model ensembles seem good to me, but more do give slightly more improvement. And models trained with different losses seem to be good at different types of events.",
    "2221276": "I think the answer depends on several factors, such as the diversity and quality of the models, the correlation between their predictions, and the weighting scheme for the ensemble."
  },
  "source": "meta"
}