{
  "id": 430218,
  "title": "How does zeroing out probability values change the result?",
  "url": "/competitions/cafa-5-protein-function-prediction/discussion/430218",
  "author_name": "",
  "post_date": "2023-08-08T18:08:24.206897100Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>If I know that a certain GO node has a probability of 0.05 let's say, almost certainly not connected to the protein, can I just not go through the children nodes and simply not include them? Does not including the children nodes' probabilities alter the final score somehow? Probably a dump question, but it's bugging me…</p>",
  "messages": [
    {
      "id": "2380706",
      "postDate": "08/08/2023 18:08:24",
      "content": "<p>If I know that a certain GO node has a probability of 0.05 let's say, almost certainly not connected to the protein, can I just not go through the children nodes and simply not include them? Does not including the children nodes' probabilities alter the final score somehow? Probably a dump question, but it's bugging me…</p>",
      "rawMarkdown": "If I know that a certain GO node has a probability of 0.05 let's say, almost certainly not connected to the protein, can I just not go through the children nodes and simply not include them? Does not including the children nodes' probabilities alter the final score somehow? Probably a dump question, but it's bugging me...",
      "votes": null
    },
    {
      "id": "2380831",
      "postDate": "08/08/2023 19:49:30",
      "content": "<p>This is a good question. I believe you are wondering about the scenario where say you get the probability for a term to be 0.5. But the probability for the parent term is 0.05. My suggestion would be to test such scenarios against a held out test dataset. See what happens when you remove the parent and child terms and see what happens when you keep them. In my case I am ignoring such propagation issues and relying entirely on the model ( not that my score is any good :) ).</p>",
      "rawMarkdown": "This is a good question. I believe you are wondering about the scenario where say you get the probability for a term to be 0.5. But the probability for the parent term is 0.05. My suggestion would be to test such scenarios against a held out test dataset. See what happens when you remove the parent and child terms and see what happens when you keep them. In my case I am ignoring such propagation issues and relying entirely on the model ( not that my score is any good :) ).",
      "votes": null
    },
    {
      "id": "2389219",
      "postDate": "08/13/2023 22:03:28",
      "content": "<p>I made a few tests to check how different filtering threshold values affect the final score. The measurements show that simply removing predictions below the given threshold lowers the final score.</p>",
      "rawMarkdown": "I made a few tests to check how different filtering threshold values affect the final score. The measurements show that simply removing predictions below the given threshold lowers the final score.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2380831,
      "author_name": "saraswat40",
      "author_url": "",
      "post_date": "08/08/2023 19:49:30",
      "content": "<p>This is a good question. I believe you are wondering about the scenario where say you get the probability for a term to be 0.5. But the probability for the parent term is 0.05. My suggestion would be to test such scenarios against a held out test dataset. See what happens when you remove the parent and child terms and see what happens when you keep them. In my case I am ignoring such propagation issues and relying entirely on the model ( not that my score is any good :) ).</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2389219,
      "author_name": "radbis",
      "author_url": "",
      "post_date": "08/13/2023 22:03:28",
      "content": "<p>I made a few tests to check how different filtering threshold values affect the final score. The measurements show that simply removing predictions below the given threshold lowers the final score.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2380706": "If I know that a certain GO node has a probability of 0.05 let's say, almost certainly not connected to the protein, can I just not go through the children nodes and simply not include them? Does not including the children nodes' probabilities alter the final score somehow? Probably a dump question, but it's bugging me...",
    "2380831": "This is a good question. I believe you are wondering about the scenario where say you get the probability for a term to be 0.5. But the probability for the parent term is 0.05. My suggestion would be to test such scenarios against a held out test dataset. See what happens when you remove the parent and child terms and see what happens when you keep them. In my case I am ignoring such propagation issues and relying entirely on the model ( not that my score is any good :) ).",
    "2389219": "I made a few tests to check how different filtering threshold values affect the final score. The measurements show that simply removing predictions below the given threshold lowers the final score."
  },
  "source": "meta"
}