Bayesian Joint Optimization for Topic Model and Clustering

摘要

Statistical clustering is the method for dividing the given samples by assumed distributions. In high dimensional problems, such as document or image clustering, the direct method is suffered from over-fitting and the curse of the dimensionality. In many cases, we firstly reduce the dimensionality, then apply the clustering algorithm. However these methods neglect the interaction among two processes. In this report, we propose the hierarchical joint distribution of Latent Dirichlet Allocation and Polya Mixture and give the parameter estimation algorithm by Gibbs sampling method. Some benchmarks show the effectiveness of the proposed method.