Robust Dirichlet Process mixtures

Non-parametric Dirichlet Process mixture (DPM) approaches for density estimation and clustering allow for automatic model selection. In this paper, we aim to develop robust DPM algorithm for clustering datasets with scatter objects, or outliers. In the developed mean-field variational inference algorithms, the auxiliary posterior distributions are factorized in a tree-structured form. In the experiments, we first show the advantage of the tree-structured factorization over the commonly-used full factorization. Then the performances of the robust DPM is evaluated using controlled experiment settings. Finally, the developed robust DPM is applied to biology datasets.