The statistics of Discrete Determinantal Point Processes

Victor-Emmanuel Brunel (ENSAE-CREST)
April 08, 2022 — 11:00 — "Salle des séminaires du L2S"

Abstract

Determinantal point processes are a family of probability distributions over subsets of a given ground set. They are very important in integrable probability, e.g., in the study of random matrices. In statistics and machine learning, they have been used to model random selections over finite dictionaries. In some cases, their most attractive feature is a property called negative association, which, intuitively, enforces diversity within the randomly selected items. In this talk, I will give a formal definition of these objects and discuss somestatistical problems such as identifiability and estimation.

Biography

I am currently a professor of statistics at ENSAE. My research interests are mostly centered around the interactions between data and geometry. In the past, I have worked on stochastic geometry - I studied random polytopes and their applications to statistics, and I am currently working on inference based on Riemannian data. I have also worked on various other topics, including determinantal point processes, robust and private inference, generative models, etc.