{
  "id": 309106,
  "title": "Tips on dealing with Class Imbalance!💯",
  "url": "/competitions/birdclef-2022/discussion/309106",
  "author_name": "Alex Teboul",
  "post_date": "2022-02-21T21:44:14.669000",
  "votes": 19,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Making this thread to discuss strategies to deal with class imbalance. The imbalance is pretty extreme and I'm wondering if anyone has some unique oversampling, augmentation, or modeling strategy that is optimal for this type of audio data? Or has anyone found/cleaned a set of minority class recordings?</p>\n<p><a href=\"https://postimg.cc/5Y79wvFj\" target=\"_blank\"><img src=\"https://i.postimg.cc/nrZsfGFK/Screen-Shot-2022-02-21-at-4-21-17-PM.png\" alt=\"Screen-Shot-2022-02-21-at-4-21-17-PM.png\"></a></p>\n<p><strong>Tip - Geographic based models w/Domain Expertise</strong></p>\n<ul>\n<li>One of the goals of this competition is to be able to reliably classify endangered or rare bird species that exist only in certain places like an Island in Hawaii. </li>\n<li>Take the Kiwikiu (Maui Parrotbill) for example which is only found within 50 square kilometres of mesic and wet forests at 1,200–2,150 metres on the windward slopes of Haleakalā. </li>\n<li>We only have 1 recording of this bird!</li>\n<li>Makes sense to have models more sensitive to this bird species and environmentally appropriate sounds when looking to identify this sound!</li>\n<li>I believe the hosts would see improved model performance if they made available a test_metadata.csv that included [latitude, longitude] for the sounds we will be submitting on. That way models could actually use the metadata..</li>\n</ul>\n<p><a href=\"https://postimg.cc/r0mRXq10\" target=\"_blank\"><img src=\"https://i.postimg.cc/B65TF6tg/2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg\" alt=\"2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg\"></a></p>\n<p>Will add more tips.</p>",
  "messages": [
    {
      "id": 1700352,
      "postDate": "2022-02-21T21:44:14.670Z",
      "content": "<p>Making this thread to discuss strategies to deal with class imbalance. The imbalance is pretty extreme and I'm wondering if anyone has some unique oversampling, augmentation, or modeling strategy that is optimal for this type of audio data? Or has anyone found/cleaned a set of minority class recordings?</p>\n<p><a href=\"https://postimg.cc/5Y79wvFj\" target=\"_blank\"><img src=\"https://i.postimg.cc/nrZsfGFK/Screen-Shot-2022-02-21-at-4-21-17-PM.png\" alt=\"Screen-Shot-2022-02-21-at-4-21-17-PM.png\"></a></p>\n<p><strong>Tip - Geographic based models w/Domain Expertise</strong></p>\n<ul>\n<li>One of the goals of this competition is to be able to reliably classify endangered or rare bird species that exist only in certain places like an Island in Hawaii. </li>\n<li>Take the Kiwikiu (Maui Parrotbill) for example which is only found within 50 square kilometres of mesic and wet forests at 1,200–2,150 metres on the windward slopes of Haleakalā. </li>\n<li>We only have 1 recording of this bird!</li>\n<li>Makes sense to have models more sensitive to this bird species and environmentally appropriate sounds when looking to identify this sound!</li>\n<li>I believe the hosts would see improved model performance if they made available a test_metadata.csv that included [latitude, longitude] for the sounds we will be submitting on. That way models could actually use the metadata..</li>\n</ul>\n<p><a href=\"https://postimg.cc/r0mRXq10\" target=\"_blank\"><img src=\"https://i.postimg.cc/B65TF6tg/2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg\" alt=\"2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg\"></a></p>\n<p>Will add more tips.</p>",
      "rawMarkdown": "Making this thread to discuss strategies to deal with class imbalance. The imbalance is pretty extreme and I'm wondering if anyone has some unique oversampling, augmentation, or modeling strategy that is optimal for this type of audio data? Or has anyone found/cleaned a set of minority class recordings?\n\n[![Screen-Shot-2022-02-21-at-4-21-17-PM.png](https://i.postimg.cc/nrZsfGFK/Screen-Shot-2022-02-21-at-4-21-17-PM.png)](https://postimg.cc/5Y79wvFj)\n\n\n**Tip - Geographic based models w/Domain Expertise**\n - One of the goals of this competition is to be able to reliably classify endangered or rare bird species that exist only in certain places like an Island in Hawaii. \n - Take the Kiwikiu (Maui Parrotbill) for example which is only found within 50 square kilometres of mesic and wet forests at 1,200–2,150 metres on the windward slopes of Haleakalā. \n - We only have 1 recording of this bird!\n - Makes sense to have models more sensitive to this bird species and environmentally appropriate sounds when looking to identify this sound!\n - I believe the hosts would see improved model performance if they made available a test_metadata.csv that included [latitude, longitude] for the sounds we will be submitting on. That way models could actually use the metadata..\n\n[![2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg](https://i.postimg.cc/B65TF6tg/2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg)](https://postimg.cc/r0mRXq10)\n\nWill add more tips.",
      "votes": 19
    },
    {
      "id": 1752711,
      "postDate": "2022-04-12T04:27:44.900Z",
      "content": "<p>Insightful!</p>",
      "rawMarkdown": "Insightful!"
    },
    {
      "id": 1752092,
      "postDate": "2022-04-11T11:50:42.057Z",
      "content": "<p>i am trying using the focal loss but I think it doesn't have much difference between it and regular categorical cross-entropy loss</p>",
      "rawMarkdown": "i am trying using the focal loss but I think it doesn't have much difference between it and regular categorical cross-entropy loss"
    },
    {
      "id": 1735710,
      "postDate": "2022-03-26T14:39:17.197Z",
      "content": "<p>Appreciate!</p>",
      "rawMarkdown": "Appreciate!"
    },
    {
      "id": 1713274,
      "postDate": "2022-03-05T20:04:30.710Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true
    },
    {
      "id": 1700572,
      "postDate": "2022-02-22T05:18:41.700Z",
      "content": "<p>Thanks, for sharing 👍</p>",
      "rawMarkdown": "Thanks, for sharing 👍",
      "votes": 1
    },
    {
      "id": 1752469,
      "postDate": "2022-04-11T18:47:45.313Z",
      "content": "<p>Thank you for sharing the insights.</p>",
      "rawMarkdown": "Thank you for sharing the insights."
    }
  ],
  "comments": [
    {
      "id": 1752711,
      "author_name": "Viswanath",
      "author_url": "",
      "post_date": "2022-04-12T04:27:44.900000",
      "content": "<p>Insightful!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1752092,
      "author_name": "Ailinz",
      "author_url": "",
      "post_date": "2022-04-11T11:50:42.057000",
      "content": "<p>i am trying using the focal loss but I think it doesn't have much difference between it and regular categorical cross-entropy loss</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1735710,
      "author_name": "Xenas",
      "author_url": "",
      "post_date": "2022-03-26T14:39:17.197000",
      "content": "<p>Appreciate!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1713274,
      "author_name": "",
      "author_url": "",
      "post_date": "2022-03-05T20:04:30.710000",
      "content": "",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1700572,
      "author_name": "Ravi_kr",
      "author_url": "",
      "post_date": "2022-02-22T05:18:41.700000",
      "content": "<p>Thanks, for sharing 👍</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1752469,
      "author_name": "Rashmita Karak",
      "author_url": "",
      "post_date": "2022-04-11T18:47:45.313000",
      "content": "<p>Thank you for sharing the insights.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1700352": "Making this thread to discuss strategies to deal with class imbalance. The imbalance is pretty extreme and I'm wondering if anyone has some unique oversampling, augmentation, or modeling strategy that is optimal for this type of audio data? Or has anyone found/cleaned a set of minority class recordings?\n\n[![Screen-Shot-2022-02-21-at-4-21-17-PM.png](https://i.postimg.cc/nrZsfGFK/Screen-Shot-2022-02-21-at-4-21-17-PM.png)](https://postimg.cc/5Y79wvFj)\n\n\n**Tip - Geographic based models w/Domain Expertise**\n - One of the goals of this competition is to be able to reliably classify endangered or rare bird species that exist only in certain places like an Island in Hawaii. \n - Take the Kiwikiu (Maui Parrotbill) for example which is only found within 50 square kilometres of mesic and wet forests at 1,200–2,150 metres on the windward slopes of Haleakalā. \n - We only have 1 recording of this bird!\n - Makes sense to have models more sensitive to this bird species and environmentally appropriate sounds when looking to identify this sound!\n - I believe the hosts would see improved model performance if they made available a test_metadata.csv that included [latitude, longitude] for the sounds we will be submitting on. That way models could actually use the metadata..\n\n[![2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg](https://i.postimg.cc/B65TF6tg/2560px-Kiwikiu-perched-in-the-Waikamoi-Forest-Preserve.jpg)](https://postimg.cc/r0mRXq10)\n\nWill add more tips.",
    "1752711": "Insightful!",
    "1752092": "i am trying using the focal loss but I think it doesn't have much difference between it and regular categorical cross-entropy loss",
    "1735710": "Appreciate!",
    "1713274": "",
    "1700572": "Thanks, for sharing 👍",
    "1752469": "Thank you for sharing the insights."
  }
}