A DCA Based Algorithm for Feature Selection in Model-Based Clustering

Abstract

Gaussian Mixture Models (GMM) is a model-based clustering approach which has been used in many applications thanks to its flexibility and effectiveness. However, in high dimension data, GMM based clustering lost its advantages due to over-parameterization and noise features. To deal with this issue, we incorporate feature selection into GMM clustering. For the first time, a non-convex sparse inducing regularization is considered for feature selection in GMM clustering. The resulting optimization problem is nonconvex for which we develop a DCA (Difference of Convex functions Algorithm) to solve. Numerical experiments on several benchmark and synthetic datasets illustrate the efficiency of our algorithm and its superiority over an EM method for solving the GMM clustering using regularization.