Positive and Unlabelled data classifier I have few customers and am supposed to find 'similar' people from a pool of customers and non-customers, what approach would you follow to solve such a problem? The key point here is that non-customer labels are not provided. I have heard of One-class classifier and Positive Unlabelled Classifier, my question is more on the lines of the caveats of using a standard classifier like Xgboost to train over this dataset by using a random set from the pool as 'non-customers'. Would this be a fair approach?
 A: I'm not sure if a "normal" classification model can help you with this problem.  You aren't really trying to predict a label.  I suppose you could try to build a classifier than predicts similar or not-similar.  But you only have labelled data for the positive class, so you're stuck with one-class classifiers.

caveats of using a standard classifier like Xgboost to train over this dataset by using a random set from the pool as 'non-customers'

So you would take a sample from the non-customer pool and label these people as not-similar?  You could try it, but the labels are going to be noisy.  Some of the people in the non-customer pool may, in fact, be similar to your seed customers.
An alternative approach is to treat this more like a clustering problem.  You have a "seed" cluster with a number of profiles.  What you need is a similarity measure between the "seed" customers and the larger pool of people (customers and non-customers).  You could threshold the similarity metric to find similar people.
