Flexible and Robust Bayesian Classification by Finite Mixture Models

Abstract. The regularized Mahalanobis distance is proposed in the framework of finite mixture models to avoid commonly faced numerical difficulties encountered with EM. Its principle is applied to Gaussian and Student-t mixtures, resulting in reliable density estimates, the model complexity being kept low. Besides, the regularized models are robust to various noise types. Finally, it is shown that the quality of the associated Bayesian classification is near optimal on Ripley's synthetic data set.

• Type: Conference
• Year: 2004
• Where: ESANN
• Authors: Cédric Archambeau, Frédéric Vrins, Michel Verleysen
• Keywords: ESANN 2004, ESANN 2007, Finite Mixture Models, Regularized Mahalanobis Distance, Regularized Models