Model selection for mixture model via integrated nested Laplace approximation

Abstract

To cluster or partition data/signal, expectation-and-maximisation or variational approximation with a mixture model (MM), which is a parametric probability density function represented as a weighted sum of (K) over cap densities, is often used. However, model selection to find the underlying (K) over cap is one of the key concerns in MM clustering, since the desired clusters can be obtained only when (K) over cap is known. A new model selection algorithm to explore (K) over cap in a Bayesian framework is proposed. The proposed algorithm builds the density of the model order which information criterion such as AIC and BIC or other heuristic algorithms basically fail to reconstruct. In addition, this algorithm reconstructs the density quickly as compared with the time-consuming Monte Carlo simulation using integrated nested Laplace approximation.