A Liouville-based approach for discrete data categorization

Abstract

In this paper, we describe a learning approach based on the smoothing of multinomial estimates using Beta-Liouville distributions. Like the Dirichlet, the Beta-Liouville is conjugate to the multinomial. It has, however, an important advantage which is its more general covariance matrix. Empirical results indicate that the proposed approach outperforms previous smoothing techniques based mainly on the Dirichlet distribution. © 2011 Springer-Verlag.