Learning, Large Scale Inference, and Temporal Modeling of Determinantal Point Processes

Determinantal Point Processes (DPPs) are random point processes well-suited for modelling repulsion. In discrete settings, DPPs are a natural model for subset selection problems where diversity is desired. For example, they can be used to select relevant but diverse sets of text or image search results. Among many remarkable properties, they offer tractable algorithms for exact inference, including computing marginals, computing certain conditional probabilities, and sampling. In this thesis, we provide four main contributions that enable DPPs to be used in more general settings. First, we develop algorithms to sample from approximate discrete DPPs in settings where we need to select a diverse subset from a large amount of items. Second, we extend this idea to continuous spaces where we develop approximate algorithms to sample from continuous DPPs, yielding a method to select point configurations that tend to be overly-dispersed. Our third contribution is in developing robust algorithms to learn the parameters of the DPP kernels, which is previously thought to be a difficult, open problem. Finally, we develop a temporal extension for discrete DPPs, where we model sequences of subsets that are not only marginally diverse but also diverse across time.