A Model Selection Approach for Flexible Dirichlet Mixtures Through Minimum Message Length

This paper introduces a model selection approach for the Flexible Dirichlet Mixture Model (FDMM), aimed at determining the optimal number of clusters while avoiding overfitting. We integrate the Minimum Message Length (MML) criterion, a Bayesian principle that jointly optimizes model complexity and data fit. Parameters are estimated via the Expectation-Maximization (EM) algorithm and initialized using the method of moments. Results on real-world datasets show that FDMM offers improved adaptability, interpretability, and robustness to structural heterogeneity compared to conventional Dirichlet-based and Gaussian mixture models.