{
  "id": 224560,
  "title": "Fractional floors allowed in submission? Good idea?",
  "url": "/competitions/indoor-location-navigation/discussion/224560",
  "author_name": "Paul Fornia",
  "post_date": "2021-03-09T02:52:42.397000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>The evaluation page states \"IMPORTANT: The integer floor used in the submission must be mapped from the char/int floors used in the dataset.\" This seems to imply that floor predictions must be integers. However, one of my submissions (and the evaluation formula) seemed to handle fractional values just fine. I.e., if model is 50/50 on F1 and F2, why not just submit 0.5?</p>\n<p>As a secondary question: Is this a good strategy? I <em>think</em> with mean-absolute-error, the expected error is equivalent whether you round to the nearest integer, or predict the expected value floor that your model spits out. </p>\n<p>In the example above, (assuming the correct answer is actually f1 or f2 each with p=0.5), the expected error is 0.5 either way. If pred=0.5, then error guaranteed to be 0.5. If you guess, then err=0 w/ p=0.5, and err=1 w/ p=0.5, so E(err) = 0.5. However, it seems weird to bias estimates by rounding. Curious to hear what folks think.</p>",
  "messages": [
    {
      "id": 1231484,
      "postDate": "2021-03-09T02:52:42.397Z",
      "content": "<p>The evaluation page states \"IMPORTANT: The integer floor used in the submission must be mapped from the char/int floors used in the dataset.\" This seems to imply that floor predictions must be integers. However, one of my submissions (and the evaluation formula) seemed to handle fractional values just fine. I.e., if model is 50/50 on F1 and F2, why not just submit 0.5?</p>\n<p>As a secondary question: Is this a good strategy? I <em>think</em> with mean-absolute-error, the expected error is equivalent whether you round to the nearest integer, or predict the expected value floor that your model spits out. </p>\n<p>In the example above, (assuming the correct answer is actually f1 or f2 each with p=0.5), the expected error is 0.5 either way. If pred=0.5, then error guaranteed to be 0.5. If you guess, then err=0 w/ p=0.5, and err=1 w/ p=0.5, so E(err) = 0.5. However, it seems weird to bias estimates by rounding. Curious to hear what folks think.</p>",
      "rawMarkdown": "The evaluation page states \"IMPORTANT: The integer floor used in the submission must be mapped from the char/int floors used in the dataset.\" This seems to imply that floor predictions must be integers. However, one of my submissions (and the evaluation formula) seemed to handle fractional values just fine. I.e., if model is 50/50 on F1 and F2, why not just submit 0.5?\n\nAs a secondary question: Is this a good strategy? I *think* with mean-absolute-error, the expected error is equivalent whether you round to the nearest integer, or predict the expected value floor that your model spits out. \n\nIn the example above, (assuming the correct answer is actually f1 or f2 each with p=0.5), the expected error is 0.5 either way. If pred=0.5, then error guaranteed to be 0.5. If you guess, then err=0 w/ p=0.5, and err=1 w/ p=0.5, so E(err) = 0.5. However, it seems weird to bias estimates by rounding. Curious to hear what folks think.",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1231484": "The evaluation page states \"IMPORTANT: The integer floor used in the submission must be mapped from the char/int floors used in the dataset.\" This seems to imply that floor predictions must be integers. However, one of my submissions (and the evaluation formula) seemed to handle fractional values just fine. I.e., if model is 50/50 on F1 and F2, why not just submit 0.5?\n\nAs a secondary question: Is this a good strategy? I *think* with mean-absolute-error, the expected error is equivalent whether you round to the nearest integer, or predict the expected value floor that your model spits out. \n\nIn the example above, (assuming the correct answer is actually f1 or f2 each with p=0.5), the expected error is 0.5 either way. If pred=0.5, then error guaranteed to be 0.5. If you guess, then err=0 w/ p=0.5, and err=1 w/ p=0.5, so E(err) = 0.5. However, it seems weird to bias estimates by rounding. Curious to hear what folks think."
  }
}