Compressing Bayesian Networks: Swarm-Based Descent, Efficiency, and Posterior Accuracy

摘要

Local models in Bayesian networks (BNs) reduce space complexity, facilitate acquisition, and can improve inference efficiency. This work focuses on Non-Impeding Noisy-AND Tree (NIN-AND Tree or NAT) models whose merits include linear complexity, being based on simple causal interactions, expressiveness, and generality. We present a swarm-based constrained gradient descent for more efficient compression of BN CPTs (conditional probability tables) into NAT models. We show empirically that multiplicatively factoring NAT-modeled BNs allows significant speed up in inference for a reasonable range of sparse BN structures. We also show that such gain in efficiency only causes reasonable approximation errors in posterior marginals in NAT-modeled real world BNs.