Title: Causal Preference Elicitation

Abstract:

This study introduces causal preference elicitation, a Bayesian approach designed for expert-in-the-loop causal discovery. This framework actively solicits information regarding local edge relationships to sharpen the posterior distribution over directed acyclic graphs (DAGs). By treating any black-box observational posterior as a starting point, we represent noisy expert opinions through a three-way likelihood function that accounts for edge existence and direction. Our method employs a flexible particle approximation for posterior inference, while query selection is driven by an efficient expected information gain criterion applied to the expert’s categorical responses. Evaluations conducted on synthetic graphs, protein signaling datasets, and a human gene perturbation benchmark demonstrate that this approach achieves quicker posterior concentration and enhances the recovery of directed effects, even when constrained by limited query budgets.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC