Fast aggregation of Student mixture models

International audienceThis paper deals with probabilistic models, that take the form of mixtures of Student distributions. Student distributions are known to be more statistically robust than Gaussian distributions, with regard to outliers (i.e. data that cannot be reasonnably explained by any component in the mixture and that do not justifiy an extra component. Our contribution is as follows : we show how several mixtures of Student distributions may be agregated into a single mixture, without resorting to sampling. The trick is that, as is well known, a Student distribution may be expressed as an infinite mixture of Gaussians, where the variances follow a Gamma distribution