A Comprehensively Improved Hybrid Algorithm for Learning Bayesian Networks: Multiple Compound Memory Erasing

Bayesian network (BN) is a classical probability graph model. It combines probability theory with graph theory to deal with uncertainty and uses a directed acyclic graph (DAG) to represent the association between nodes. It has been successfully applied to prediction [1], risk analysis [2], semantic search, biological system modeling, and other practical fields [3].This field has two components: the structure learning of Bayesian networks and the parameter learning of Bayesian networks. The latter is based on the former, and the structure learning of Bayesian networks is often more important and complex [4]. BN can enable decisionmakers to make conditional causal inferences on uncertain behaviors with the help of the causal relationship of nodes in the network and can deduce the most powerful decision nodes that affect the results. Therefore, this paper focuses on how to generate a prediction network structure with the greatest similarity to the original network structure in a short time, rather than on whether the global score of the prediction model is higher [5]. The methods of learning Bayesian network structure (BNs) from data can generally be divided into three categories: constraint-based, score-based and search strategy, and hybrid algorithms [6]. Representative constraint-based methods mainly include grow-shrink (GS) [7], three-phase dependency analysis (TPDA) [8], PC [9], and incremental associated Markov blanket analysis (IAMB) [10]. The constraint-based methods usually make conditional independence (CI) tests between nodes (i.e.