Efficient Sampling for k-Determinantal Point Processes

Determinantal Point Processes (DPPs) provide probabilistic models over discrete sets of items that help model repulsion and diversity. Applicability of DPPs to large sets of data is, however, hindered by the expensive matrix operations involved, especially when sampling. We therefore propose a new efficient approximate two-stage sampling algorithm for discrete k-DPPs. As opposed to previous approximations, our algorithm aims at minimizing the variational distance to the original distribution. Experiments indicate that the resulting sampling algorithm works well on large data and yields more accurate samples than previous approaches.