1
$\begingroup$

Let's say I am working on handwritten digit recognition (0 to 9). I know for instance that if I use clustering then I need to look for 10 clusters. But once I have the 10 clusters,how do I identify automatically which cluster corresponds to which digit? In general, say I have some other classification task where I don't have the labels for the training data, but I know that there are, lets say, two classes. I want to build a model which can then tell me, for the test instances, which classes they belong to. But again the problem that even if I can group the train cases together, how do I know the actual label of the test instance in this unsupervised setting?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

You can't, in a completely unsupervised setting. You need something more: e.g., a training data set where you have labels for at least some of the instances, or some other information that allows you to set a label for each cluster.

How is the computer supposed to know that a circle means zero as opposed to some other number, if you don't tell it that? Obviously, it can't. In fact, in the Eastern Arabic numbering system, there's a zero-looking character that actually means the number 5, not the number 0. Obviously, a computer can't possibly know which of those two meanings is intended, if you don't tell it somehow.

Improve this answer
$\endgroup$
3
  • 1
    $\begingroup$ To be precise, you don't need a training set to assign an object to a specific class. You need classification rules that can be obtained from a training set. $\endgroup$
    – Marcin Król
    Dec 11, 2015 at 9:13
  • $\begingroup$ Coud you please elaborate? $\endgroup$
    – user3676846
    Dec 11, 2015 at 13:10
  • $\begingroup$ I had this classification task, where I was supposed to predict teachers who stayed at a particular school vs those that left the school for better opportunities.The problem was I only had data of the current teachers in the school,so I tried to ask some teachers to pretend as if they had the attributes of someone who would be likely to leave the school. Given such a setting, would it make sense to first cluster using say, k-means with k=2, then somehow determine which cluster denoted the teachers who left, and then assign labels based on these, and to further treat this as a supervised task? $\endgroup$
    – user3676846
    Dec 11, 2015 at 13:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.