Tree-Structured Clustering via the Minimum Cross Entropy Principle

摘要: We propose a new interdisciplinary approach to the tree-structured clustering problem, wherein structural constraints are imposed in order reduce classification search complexity of resulting statistical classifier. Most known methods greedy and optimize nodes tree one at time minimize local cost. By constrast, we develop joint optimization method, derived based on information-theoretic principles closely related physics. The is inspired by deterministic annealing method for unstructured clustering, which was maximum entropy inference. principle minimum cross entropy, using informative priors approximate solution while imposing constraint. As original number distinct representatives (and hence tree) grows non-heuristic fashion sequence phase transitions occur so as effective free energy Examples demonstrate considerable improvement over methods.