An improved Harris Hawks optimization for Bayesian network structure learning via genetic operators

Constructing Bayesian network structures from data is an NP-hard problem. This paper presents a novel method for Bayesian network structure learning using a discrete Harris hawks optimization algorithm, named BNC-HHO. It uses the max-min parents and children algorithm, V-structure & log-likelihood function, and neighborhood structures to limit the search space during the initialization phase. Then, the Harris hawk optimization algorithm is extended from the continuous to the discrete domain by redefining the movement strategies of hawks using genetic operators in genetic algorithm. The crossover and mutation operations in the proposed method are controlled by an adaptive crossover and mutation rate based on the X-conditional cloud. To balance the exploration and exploitation phases, a nonlinear escaping energy curve is also designed. Finally, the quality of the solution is further improved using a local optimizer. Experiments on various standard networks demonstrate that the proposed algorithm can quickly get higher structure scores and better convergence accuracy in most cases compared to other state-of-the-art algorithms. It indicates that the proposed algorithm can be used as an effective and feasible method for learning Bayesian network structures.