CD-BNN: Causal Discovery with Bayesian Neural Network

Huaxu Han, Shuliang Wang, Hanning Yuan, Sijie Ruan

Abstract: Causal discovery involves learning Directed Acyclic Graphs (DAGs) from observational data and has widespread applications in various fields. Recent advancements in the structural equation model (SEM) have successfully applied continuous optimization techniques to causal discovery. These methods introduce acyclicity constraints to tackle the challenge of exploring the exponentially large search space that arises as the number of graph nodes increases. However, these methods often rely on point estimates that fail to fully account for the inherent uncertainty present in the data. This limitation can lead to inaccurate causal graph inference. In this paper, we propose a novel method for causal discovery with Bayesian Neural Networks (CD-BNN). CD-BNN incorporates a Bayesian Neural Network to explicitly model and quantify uncertainty in the data while reducing the influence of noise through model averaging. Moreover, we explore the extraction of the final DAG from the BNN using partial derivatives. We conduct a comprehensive set of experiments on both real-world and synthetic data to evaluate the performance of our approach. The results demonstrate that our proposed method surpasses related baselines in accurately identifying causal graphs, particularly when faced with data uncertainty.