Bayesian network structure learning based on discrete artificial jellyfish search: Leveraging scoring and graphical properties

Abstract

It is challenging to obtain a credible Bayesian network (BN) from data to represent uncertain knowledge. Current swarm intelligence optimization methods lack targeted improvements based on domain-specific knowledge, which limits the learning accuracy and convergence speed. To address this gap, we propose a novel discrete artificial jellyfish search method for structure learning that leverages the scoring and graphical properties of BNs. Inspired by scoring functions and equivalence classes, a directional crossover operator is designed to efficiently narrow the crossover range. Additionally, a bidirectional search operator uses score increment guidance during mutations. By incorporating adjacency matrix series, global loop finding and deleting operators are applied to identify and eliminate all the minimalist loops simultaneously. They can avoid omitting the optimal solution. The experimental results show that the proposed algorithm outperforms the existing state-of-the-art algorithms in the scoring and convergence speed, which achieves an effective integration of group intelligence and structure learning.