Landmark-Dependent Hierarchical Beta Processfor Robust Sparse Factor Analysis

A computationally efficient landmark-dependent hierarchical beta process is developed as a prior for data with associated covariates. The landmarks define local regions in the covariate space where feature usages are likely to be similar. The landmark locations are learned, to which the data are linked through normalized kernels. The proposed model is well suited for local latent feature discovery, and adding a robustness term, it successfully separates out non-local sparse spiky components, as demonstrated in image denoising and document analysis applications.