Learning causal DAG without Acyclic Constraint

Abstract

Learning the causal structure from samples of a joint distribution is a challenging problem, since the search space of the underlying Directed Acyclic Graphs (DAGs) is combinatorial. Some recent formulations turn this problem into a continuous optimization with a structural constraint enforcing acyclicity. We propose a novel formulation that ensures the learned graphs are acyclic without adding the acyclicity constraint and therefore turn the constrained optimization problem into an unconstrained one. We compare our approach to existing continuous optimization methods on real and synthetic data, and demonstrate that our method learns appropriate DAGs more efficiently thanks to the relaxation of the acyclic constraint.