{
  "id": 12587,
  "title": "Manual vs Auto feature selection",
  "url": "/competitions/datasciencebowl/discussion/12587",
  "author_name": "",
  "post_date": "2015-02-22T20:06:06.600Z",
  "votes": null,
  "comment_count": 1,
  "views": 1279,
  "content": "<p>Is it a legal option to manually select features from the training data to be used as feature examples in a decision tree or other classifier? &nbsp;For instance, manually cropping out examples of fins, tentacles, etc. and using them as input to a classifier? &nbsp;Would this fall under the semi-supervised category? &nbsp;I'm uncertain as to how much of the classifier construction can be assembled by hand/coding vs algorithmically generated?</p>",
  "messages": [
    {
      "id": "64742",
      "postDate": "02/22/2015 20:06:06",
      "content": "<p>Is it a legal option to manually select features from the training data to be used as feature examples in a decision tree or other classifier? &nbsp;For instance, manually cropping out examples of fins, tentacles, etc. and using them as input to a classifier? &nbsp;Would this fall under the semi-supervised category? &nbsp;I'm uncertain as to how much of the classifier construction can be assembled by hand/coding vs algorithmically generated?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64743",
      "postDate": "02/22/2015 20:14:49",
      "content": "<p>Manually selecting features is okay. Manually cropping the training set to improve a&nbsp;classifier is okay. Manually cropping the test set would not be okay.</p>\n<p>Use this rule of thumb - If you were to get a totally new test set tomorrow, your method should be able to classify it with comparable performance, without manual intervention.</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 64743,
      "author_name": "wcukierski",
      "author_url": "",
      "post_date": "02/22/2015 20:14:49",
      "content": "<p>Manually selecting features is okay. Manually cropping the training set to improve a&nbsp;classifier is okay. Manually cropping the test set would not be okay.</p>\n<p>Use this rule of thumb - If you were to get a totally new test set tomorrow, your method should be able to classify it with comparable performance, without manual intervention.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "64742": "",
    "64743": ""
  },
  "source": "meta"
}