A flexible approach to high-dimensional clustering and classification using mixtures of finite mutually repulsive mixtures

We propose a novel approach for clustering that is later extended to classification. Our methods are based on an underlying determinantal point process (DPP) that generates sparsity and separation of location parameters in a mixture of finite mixture models specification. A key aspect is the assumption of a DPP conditional on the number of groups and a prior on this number, which results in a dynamic mixture of repulsive mixtures. The proposed model also involves a sparse tensor decomposition of the covariance function used to define the DPP, which is coupled with a separate mixture model employed to handle the intra-cluster correlation structure. Further computational enhancement in the high-dimensional case comes from exploiting a mapping from Euclidean space to the space of (valid) correlation matrices. A suitable posterior simulation scheme involving a modified telescopic sampler is developed. The resulting approach is seen to perform satisfactorily in simulated and real data examples.