Clustering

We introduce a Bayesian Fisher-EM (BFEM) algorithm for the discriminative latent mixture model, modeling data as a mixture of Gaussians in a low-dimensional discriminative subspace (in Fisher's linear discriminant analysis sense). We demonstrate the interest of the latter in two thorough simulations settings, and propose an illustration on the unsupervised problem of image denoising with Gaussian mixture models.

Dicrete latent variable models (DLVM) are a unified framework for generative clustering models including mixture models and block models. We propose a new algorithm to greedily maximise the integrated likelihood criterion with respect to a partition in any DLVM where this quantity is analytic.

PhD. Thesis

We propose a new model for the clustering of count data that integrates mixture modeling with the dimension reduction aspect of topic modeling.