The Power of Collaboration: Learning Large Bayesian Networks at Scale

We present a novel approach to learn the structure of large Bayesian Networks (BNs) of bounded treewidth. Our approach combines the complementary strengths of heuristic and MaxSAT-based methods. Both methods operate in parallel and cooperate in a mutually advantageous way to improve the quality of the learned BN structures. Our work utilizes an approach proposed by Peruvemba Ramaswamy and Szeider (AAAI'21, NeurIPS'21, UAI'22), which continually applies a MaxSAT-based algorithm to a BN obtained from heuristic search. We create a continuous and dynamic refinement process by allowing the heuristic and the MaxSAT-based technique to cooperate and repeatedly share their work. Our rigorous experiments show that the collaboration between the two methods is a powerful strategy for increasing the score of the learned BN structure. Moreover, the rate at which the score increases is significantly higher than that of the individual methods by themselves. Our results provide a strong argument for combining complementary approaches to learn treewidth-bounded BNs in a collaborative way.