I have a data set of profiles that describes a set of people and a binary classification describing when some of these people have been grouped together in the past which I am assuming implies similarity.
I would like to classify every person in this data set as either similar or dissimilar.
One option is to take the features of profile A and profile B, calculate the distance between each of them individually, and then use these distances to predict similarity or dissimilarity.
Another option is to use the features of profile A and profile B directly to predict similarity or dissimilarity.
Once I've decided on feature preprocessing this problem becomes a "Positive Unlabeled" or PU problem that I may be able to deal with using domain knowledge (that is I know which profiles will never appear together and can thus label them negative).
I'm looking for advice on how to approach defining the features in this problem.