In this paper, we compare three heuristic methods for choosing which of a set of features to use to represent a domain of stimuli when we know the categories to which those stimuli belong. Our methods are based on three measures of category differentiation: cue validity, category validity, and their product, collocation. In a comparison of their ability to predict human similarity ratings in the Leuven Natural Concept Database, we find collocation to have the best performance, suggesting people use both cue and category validities in choosing which features to represent.