Speaker: Ryan Thompson, University of Technology Sydney (UTS), Australia
Title: ProDAG: Projection-Induced Variational Inference for Directed Acyclic Graphs
Abstract: Directed acyclic graph (DAG) learning is a rapidly evolving area with significant implications for causal inference. Despite recent advances, learning a single DAG from data—let alone quantifying graph uncertainty—remains statistically and computationally challenging. To address this issue, we propose a Bayesian variational inference framework based on novel distributions that have support directly on the space of DAGs. These distributions, which serve as our prior and variational posterior, are induced by a projection operation that maps a continuous distribution onto the space of acyclic adjacency matrices. Although this projection is a combinatorial optimization problem, it is solvable at scale using recently developed continuous characterizations of acyclicity. We demonstrate that our method, ProDAG, delivers higher quality Bayesian inference than existing state-of-the-art alternatives.
Website: ryan-thompson.github.io