, Volume 50, Issue 1-2, pp 73-94

Mixtures of Factor Analysers. Bayesian Estimation and Inference by Stochastic Simulation


Factor Analysis (FA) is a well established probabilistic approach to unsupervised learning for complex systems involving correlated variables in high-dimensional spaces. FA aims principally to reduce the dimensionality of the data by projecting high-dimensional vectors on to lower-dimensional spaces. However, because of its inherent linearity, the generic FA model is essentially unable to capture data complexity when the input space is nonhomogeneous. A finite Mixture of Factor Analysers (MFA) is a globally nonlinear and therefore more flexible extension of the basic FA model that overcomes the above limitation by combining the local factor analysers of each cluster of the heterogeneous input space. The structure of the MFA model offers the potential to model the density of high-dimensional observations adequately while also allowing both clustering and local dimensionality reduction. Many aspects of the MFA model have recently come under close scrutiny, from both the likelihood-based and the Bayesian perspectives. In this paper, we adopt a Bayesian approach, and more specifically a treatment that bases estimation and inference on the stochastic simulation of the posterior distributions of interest. We first treat the case where the number of mixture components and the number of common factors are known and fixed, and we derive an efficient Markov Chain Monte Carlo (MCMC) algorithm based on Data Augmentation to perform inference and estimation. We also consider the more general setting where there is uncertainty about the dimensionalities of the latent spaces (number of mixture components and number of common factors unknown), and we estimate the complexity of the model by using the sample paths of an ergodic Markov chain obtained through the simulation of a continuous-time stochastic birth-and-death point process. The main strengths of our algorithms are that they are both efficient (our algorithms are all based on familiar and standard distributions that are easy to sample from, and many characteristics of interest are by-products of the same process) and easy to interpret. Moreover, they are straightforward to implement and offer the possibility of assessing the goodness of the results obtained. Experimental results on both artificial and real data reveal that our approach performs well, and can therefore be envisaged as an alternative to the other approaches used for this model.