{
  "id": 182551,
  "title": "How to blend model with different best thread",
  "url": "/competitions/birdsong-recognition/discussion/182551",
  "author_name": "",
  "post_date": "2020-09-13T10:26:08.523431400Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I try to blend some models, but their threads are very different. Is that dividing each thread before blending a good method?</p>",
  "messages": [
    {
      "id": "1008703",
      "postDate": "09/13/2020 10:26:08",
      "content": "<p>I try to blend some models, but their threads are very different. Is that dividing each thread before blending a good method?</p>",
      "rawMarkdown": "I try to blend some models, but their threads are very different. Is that dividing each thread before blending a good method?",
      "votes": null
    },
    {
      "id": "1009218",
      "postDate": "09/13/2020 18:35:59",
      "content": "<p>That's a good question. For sure, your threshold depends not only from your model but from a validation dataset - try to calculate locally your threshold for a blend. But I and other participants still recommend making models threshold agnostic. That's right and cool in theory but to be true my optimal LB threshold is not 0.5. We'll see</p>",
      "rawMarkdown": "That's a good question. For sure, your threshold depends not only from your model but from a validation dataset - try to calculate locally your threshold for a blend. But I and other participants still recommend making models threshold agnostic. That's right and cool in theory but to be true my optimal LB threshold is not 0.5. We'll see",
      "votes": null
    },
    {
      "id": "1009502",
      "postDate": "09/14/2020 03:33:08",
      "content": "<p>Thanks for answering.</p>",
      "rawMarkdown": "Thanks for answering.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1009218,
      "author_name": "yaroshevskiy",
      "author_url": "",
      "post_date": "09/13/2020 18:35:59",
      "content": "<p>That's a good question. For sure, your threshold depends not only from your model but from a validation dataset - try to calculate locally your threshold for a blend. But I and other participants still recommend making models threshold agnostic. That's right and cool in theory but to be true my optimal LB threshold is not 0.5. We'll see</p>",
      "votes": null,
      "replies": [
        {
          "id": 1009502,
          "author_name": "karlyukang",
          "author_url": "",
          "post_date": "09/14/2020 03:33:08",
          "content": "<p>Thanks for answering.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1008703": "I try to blend some models, but their threads are very different. Is that dividing each thread before blending a good method?",
    "1009218": "That's a good question. For sure, your threshold depends not only from your model but from a validation dataset - try to calculate locally your threshold for a blend. But I and other participants still recommend making models threshold agnostic. That's right and cool in theory but to be true my optimal LB threshold is not 0.5. We'll see",
    "1009502": "Thanks for answering."
  },
  "source": "meta"
}