Fitting mixtures of von Mises distributions via noise contrastive estimation

Directional distributions requires the evaluation of complicated normalizing constants, even for the univariate von Mises. For this reason, maximum likelihood estimation methods are often difficult to apply in practice. To address this issue, we present an approach based on Noise Contrastive Estimation (NCE), a statistical learning technique used for parameter estimation in non-normalized statistical models. In NCE, the estimation problem is reformulated as a binary classification task. In this paper, we focus on fitting mixtures of von Mises distributions, with particular emphasis on toroidal data. Our application to real data, in which we compare several estimation methods, suggests that NCE is a promising alternative for parameter inference in finite mixtures of directional distributions.