{
  "id": 398898,
  "title": "removing the lowest predictions and normalises the rest improves score",
  "url": "/competitions/birdclef-2023/discussion/398898",
  "author_name": "",
  "post_date": "2023-04-01T09:53:35.331128Z",
  "votes": 7,
  "comment_count": 4,
  "views": 0,
  "content": "<p>see here: <a href=\"https://www.kaggle.com/code/domdejonge/remove-lowest-probability-predictions?scriptVersionId=124166248\" target=\"_blank\">https://www.kaggle.com/code/domdejonge/remove-lowest-probability-predictions?scriptVersionId=124166248</a></p>\n<p>logically it makes sense but I will experiment with other normalisation techniques (eg exponential)</p>",
  "messages": [
    {
      "id": "2205150",
      "postDate": "04/01/2023 09:53:35",
      "content": "<p>see here: <a href=\"https://www.kaggle.com/code/domdejonge/remove-lowest-probability-predictions?scriptVersionId=124166248\" target=\"_blank\">https://www.kaggle.com/code/domdejonge/remove-lowest-probability-predictions?scriptVersionId=124166248</a></p>\n<p>logically it makes sense but I will experiment with other normalisation techniques (eg exponential)</p>",
      "rawMarkdown": "see here: https://www.kaggle.com/code/domdejonge/remove-lowest-probability-predictions?scriptVersionId=124166248\n\nlogically it makes sense but I will experiment with other normalisation techniques (eg exponential)",
      "votes": null
    },
    {
      "id": "2205208",
      "postDate": "04/01/2023 10:40:39",
      "content": "<p>I think it's understandable since original train files feature presence of bird on the scale of present/not present (1/0). So it makes sense to round values which are very close to 0 to 0. I think the same thing can be done with values very close to 1 (like &gt; 0.9999) - they can be rounded to 1.</p>",
      "rawMarkdown": "I think it's understandable since original train files feature presence of bird on the scale of present/not present (1/0). So it makes sense to round values which are very close to 0 to 0. I think the same thing can be done with values very close to 1 (like > 0.9999) - they can be rounded to 1.",
      "votes": null
    },
    {
      "id": "2205217",
      "postDate": "04/01/2023 10:47:34",
      "content": "<p>yes, but i dont think there are any examples of this happening as there would need to be over 100 classes with under 0.0001 for that to be the case</p>",
      "rawMarkdown": "yes, but i dont think there are any examples of this happening as there would need to be over 100 classes with under 0.0001 for that to be the case",
      "votes": null
    },
    {
      "id": "2205433",
      "postDate": "04/01/2023 14:47:04",
      "content": "<p>Why would you normalize the predictions per time window. Multiple birds can occur in one window, so it's a multi-label problem.</p>",
      "rawMarkdown": "Why would you normalize the predictions per time window. Multiple birds can occur in one window, so it's a multi-label problem.",
      "votes": null
    },
    {
      "id": "2205904",
      "postDate": "04/02/2023 05:02:25",
      "content": "<p>Is it possible that in some time window there is no bird? In that case, does the normalization still make sense?</p>",
      "rawMarkdown": "Is it possible that in some time window there is no bird? In that case, does the normalization still make sense?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2205208,
      "author_name": "atamazian",
      "author_url": "",
      "post_date": "04/01/2023 10:40:39",
      "content": "<p>I think it's understandable since original train files feature presence of bird on the scale of present/not present (1/0). So it makes sense to round values which are very close to 0 to 0. I think the same thing can be done with values very close to 1 (like &gt; 0.9999) - they can be rounded to 1.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2205217,
          "author_name": "domdejonge",
          "author_url": "",
          "post_date": "04/01/2023 10:47:34",
          "content": "<p>yes, but i dont think there are any examples of this happening as there would need to be over 100 classes with under 0.0001 for that to be the case</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2205433,
      "author_name": "group16",
      "author_url": "",
      "post_date": "04/01/2023 14:47:04",
      "content": "<p>Why would you normalize the predictions per time window. Multiple birds can occur in one window, so it's a multi-label problem.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2205904,
      "author_name": "aphysict",
      "author_url": "",
      "post_date": "04/02/2023 05:02:25",
      "content": "<p>Is it possible that in some time window there is no bird? In that case, does the normalization still make sense?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2205150": "see here: https://www.kaggle.com/code/domdejonge/remove-lowest-probability-predictions?scriptVersionId=124166248\n\nlogically it makes sense but I will experiment with other normalisation techniques (eg exponential)",
    "2205208": "I think it's understandable since original train files feature presence of bird on the scale of present/not present (1/0). So it makes sense to round values which are very close to 0 to 0. I think the same thing can be done with values very close to 1 (like > 0.9999) - they can be rounded to 1.",
    "2205217": "yes, but i dont think there are any examples of this happening as there would need to be over 100 classes with under 0.0001 for that to be the case",
    "2205433": "Why would you normalize the predictions per time window. Multiple birds can occur in one window, so it's a multi-label problem.",
    "2205904": "Is it possible that in some time window there is no bird? In that case, does the normalization still make sense?"
  },
  "source": "meta"
}