I want to do supervised learning. I have a labeled dataset. Each label is normally a real number between [0,1] corresponding to a location of a certain object in the data, or NAN if the object is not present in the data.

Is there a "recommended way" to convert my labels into valid labels?

My ideas are:

  1. Train two different classifiers, one for existence of object, and one for its location
  2. Train one classifier, but with two outputs, one binary for existence, and one float for location
  3. Relabel all NANs as -1, and train one classifier that can be -1 or in [0,1]
  • $\begingroup$ ignore approach 1 - it's less automatic than other two, for approach 3 you can set NAN to -3 or lower to be sure you will not make mistake, after all machine learning is experimental... $\endgroup$
    – quester
    Oct 8 '19 at 13:19
  • 1
    $\begingroup$ The first option seems to be the most straightforward one, but evaluating which is more approapriate should depend on the problem. A similar discussion can be found here: stats.stackexchange.com/questions/351673/… $\endgroup$ Oct 8 '19 at 13:19
  • $\begingroup$ @quester Sorry if I was unclear. By "recommended" I mean links to examples of solutions to same or related problems that are already well-studied by their authors. I would like to avoid experimenting with parameters that likely have already been investigated by people more professional in the field than myself $\endgroup$ Oct 8 '19 at 14:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.