{
  "id": 307775,
  "title": "Triplets/Siamese/classification or Any other",
  "url": "/competitions/happy-whale-and-dolphin/discussion/307775",
  "author_name": "",
  "post_date": "2022-02-15T15:32:26.275833400Z",
  "votes": 11,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Opening this discussion for best approach in this competitions</p>\n<p>Triplets:<br>\n1 Given a individual id , you look for one similar instance and other dissimilar instance from same species </p>\n<p>Siamese</p>\n<p>1) Given a individual id , you look for dissimilar id or similar id and learn similarity /dissimilarity </p>\n<p>Classification :</p>\n<p>1) Simple take distinct class and go..</p>\n<p>So far public notebook are leaning towards  classification approach owing to its high score. <br>\nBut could that high score be result of easy species present in public set ?so in that case can classification work well for rare  individual id classification . Remember   12k individual id  have got just 1 or 2 instances . May be public lb consists of other 3k species  which are abundant in quantity .</p>\n<p>Similar  can be constraint for  Triplets/Siamese . for many species we can learn only dissimilarity well .<br>\nIn that case what can be best plausible approach . I open this topic so people can focus on right approach rather following  one single public notebook and get overhauled in silver/bronze zone. </p>",
  "messages": [
    {
      "id": "1691748",
      "postDate": "02/15/2022 15:32:26",
      "content": "<p>Opening this discussion for best approach in this competitions</p>\n<p>Triplets:<br>\n1 Given a individual id , you look for one similar instance and other dissimilar instance from same species </p>\n<p>Siamese</p>\n<p>1) Given a individual id , you look for dissimilar id or similar id and learn similarity /dissimilarity </p>\n<p>Classification :</p>\n<p>1) Simple take distinct class and go..</p>\n<p>So far public notebook are leaning towards  classification approach owing to its high score. <br>\nBut could that high score be result of easy species present in public set ?so in that case can classification work well for rare  individual id classification . Remember   12k individual id  have got just 1 or 2 instances . May be public lb consists of other 3k species  which are abundant in quantity .</p>\n<p>Similar  can be constraint for  Triplets/Siamese . for many species we can learn only dissimilarity well .<br>\nIn that case what can be best plausible approach . I open this topic so people can focus on right approach rather following  one single public notebook and get overhauled in silver/bronze zone. </p>",
      "rawMarkdown": "Opening this discussion for best approach in this competitions\n\nTriplets:\n1 Given a individual id , you look for one similar instance and other dissimilar instance from same species \n\nSiamese\n\n1) Given a individual id , you look for dissimilar id or similar id and learn similarity /dissimilarity \n\n\nClassification :\n\n1) Simple take distinct class and go..\n\n\nSo far public notebook are leaning towards  classification approach owing to its high score. \nBut could that high score be result of easy species present in public set ?so in that case can classification work well for rare  individual id classification . Remember   12k individual id  have got just 1 or 2 instances . May be public lb consists of other 3k species  which are abundant in quantity .\n\nSimilar  can be constraint for  Triplets/Siamese . for many species we can learn only dissimilarity well .\nIn that case what can be best plausible approach . I open this topic so people can focus on right approach rather following  one single public notebook and get overhauled in silver/bronze zone.",
      "votes": null
    },
    {
      "id": "1695014",
      "postDate": "02/17/2022 22:40:57",
      "content": "<p>Hi, how would you make contrastive loss working with such a large fraction of IDs that occurs only once? Do you think similar species as positive anchors would provide meaningful information about the IDs? </p>",
      "rawMarkdown": "Hi, how would you make contrastive loss working with such a large fraction of IDs that occurs only once? Do you think similar species as positive anchors would provide meaningful information about the IDs?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1695014,
      "author_name": "tlipss",
      "author_url": "",
      "post_date": "02/17/2022 22:40:57",
      "content": "<p>Hi, how would you make contrastive loss working with such a large fraction of IDs that occurs only once? Do you think similar species as positive anchors would provide meaningful information about the IDs? </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1691748": "Opening this discussion for best approach in this competitions\n\nTriplets:\n1 Given a individual id , you look for one similar instance and other dissimilar instance from same species \n\nSiamese\n\n1) Given a individual id , you look for dissimilar id or similar id and learn similarity /dissimilarity \n\n\nClassification :\n\n1) Simple take distinct class and go..\n\n\nSo far public notebook are leaning towards  classification approach owing to its high score. \nBut could that high score be result of easy species present in public set ?so in that case can classification work well for rare  individual id classification . Remember   12k individual id  have got just 1 or 2 instances . May be public lb consists of other 3k species  which are abundant in quantity .\n\nSimilar  can be constraint for  Triplets/Siamese . for many species we can learn only dissimilarity well .\nIn that case what can be best plausible approach . I open this topic so people can focus on right approach rather following  one single public notebook and get overhauled in silver/bronze zone.",
    "1695014": "Hi, how would you make contrastive loss working with such a large fraction of IDs that occurs only once? Do you think similar species as positive anchors would provide meaningful information about the IDs?"
  },
  "source": "meta"
}