A0579
Title: Dynamic mixture of finite mixtures of factor analyzers
Authors: Sylvia Fruehwirth-Schnatter - WU Vienna University of Economics and Business (Austria)
Margarita Grushanina - Imperial College London (United Kingdom) [presenting]
Abstract: Mixtures of factor analyzers represent a popular tool for finding structure in data. While in many applications, the number of clusters and latent factors within clusters is held constant, some recent models automatically infer cluster and/or factor dimensionalities. This is done by employing nonparametric priors and allowing the number of clusters and factors to potentially be infinite. MCMC estimation is performed via adaptive algorithms, where parameters associated with the redundant factors are discarded. The current work contributes to the literature by allowing automatic inference on the number of clusters and cluster-specific factors while keeping both dimensions finite. For automatic inference on the cluster structure, the dynamic mixture of finite mixtures model is employed with a prior on the number of mixture components. Automatic inference on cluster-specific factors is performed by assigning an exchangeable shrinkage process (ESP) prior, which can be interpreted as a generalized cumulative shrinkage process (CUSP) prior to the columns of the factor loading matrices. Extensive simulation studies and applications to benchmark as well as real data sets demonstrate that the model outperforms competing alternatives, in particular, based on the multiplicative gamma process prior, with respect to recovering the correct number of cluster-specific factors.
