Many researchers have considered probabilistic models of music structure representation. However, these models have yet to provide a truly generative framework, capable of elucidating the cognitive processes of induction, concept formation, or categorization. Feature centrality examines dependency networks of harmonic units within short phrases as a possible foundation for this type of inference. We have found that features of music structures do not adhere to the same feature centrality principles as artifacts or essentialized categories. We are continuing to develop this model through subject testing and computational modeling.