Bayesian Model Averaging with Diffused Priors for Model-Based Clustering Under a Cluster Forests Architecture

This paper considers a class of generative graphical models for parsimonious modeling of Gaussian mixtures and robust unsupervised learning, each assuming that the data are generated independently and identically from a finite mixture model with an extended naïve Bayes structure. To account for model uncertainty, the expectation model-averaging algorithm, which approximates the Bayesian model averaging with incomplete data, is introduced using a novel class of non-informative priors for the parameters. A Cluster Forests architecture to circumvent intractable model averaging over a large selective model space is developed. Extensive synthetic data experiments and real-world data applications show that the proposed methodology can produce clustering results of high robustness and attain good model detection performance.