Split-and-Merge Sampling Algorithm for Hamming-Mixture Models of Categorical Data

This work aims to design a Gibbs sampling algorithm for posterior Bayesian inference of a Dirichlet process mixture model based on Hamming distributed kernels, a probability measure built upon the Hamming distance. This model is employed to provide model-based clustering analysis of categorical data with no natural ordering. The proposed algorithm leverages a split-and-merge Markov chain Monte Carlo technique to address the curse of dimensionality issue and improve the search over the space of random partitions.