The work is dedicated to classification of predicate constraints that arise in supervised learning. Suppose that supplementary constraints are given in the form of predicate pair. Then the characteristic property of supervised learning is that it require efficient algorithm to find a function preserving this pair and behaving on the training set maximally si