Research On Knowledge-Driven Bayesian Network Structure Learning Algorithm

Bayesian Networks(BN)are used for reasoning and modeling uncertain relationships.BN can help us better understand and deal with uncertainty,so it is widely used in fields such as food safety,healthcare,natural language processing,and financial engineering.The task of learning the BN structure from data,called Bayesian Network Structure Learning(BNSL),has been proven to be an NP-hard problem due to the exponentially increasing search space with the increase of problem variables.Evolutionary computation-based methods have been widely applied in BNSL,and the strategy based on Genetic Algorithm(GA)has achieved good results.However,current algorithms still suffer from problems such as low precision of the discovered structures,slow convergence speed,and high computational resource consumption due to insufficient and inadequate utilization of existing information.This paper proposes using the elite knowledge information of the population evolution to drive the algorithm and using mutual information to guide the search behavior of the population.The main work of this paper is as follows:(1)A hybrid genetic algorithm based on dual elite structure is proposed for effective search of BN structures in response to the problem of poor search accuracy due to insufficient utilization of elite knowledge information by GA.An improved elite mechanism is introduced to select elite individuals reasonably.A novel dual elite structure drawn from the elites is proposed to represent the common independency and dependency relationships simultaneously.The dual elite structure can help search in more promising regions but also help reduce the search space.Furthermore,the relative and tendency structures are drawn from the elites to use the structural information comprehensively.Diverse strategies are designed to drive the structural evolution among the three structures.The elite structure diversity is controlled within a healthy interval during the structural evolution to maintain the balance between exploration and exploitation.The proposed algorithm has been tested on nine Bayesian network datasets and compared with both traditional BNSL algorithms and state-of-the-art algorithms.The experimental findings suggest that the proposed algorithm performs better than the compared algorithms in terms of search accuracy.(2)A genetic algorithm based on quadratic mutual information guidance is proposed,taking into account the unique advantages of population in evolutionary algorithms and the limitations of mutual information in evolutionary algorithm applications.The principle of mutual information and its inspiration in this field are combined to propose this new algorithm.A strategy for removing redundant nodes based on mutual information is proposed to address the maximum parent limit of nodes in BN.In addition,according to the guidance of mutual information,the direction of variation of individuals in the variation shows a preference for mutual information and protects the edges with high mutual information values.The proposed algorithm was tested on nine BN datasets,and the experimental results validated the effectiveness of the quadratic mutual information improvement strategy in BNSL.By combining the advantages of the above improvements,the algorithm showed significant improvements in convergence behavior and accuracy of the final structure.Especially,there was a relatively significant performance improvement in the ultra-large weakly connected causal network and outperformed various compared algorithms.Overall,this paper improves the performance and efficiency of the BNSL algorithm by optimizing it with both elite knowledge-driven strategies and mutual information strategies.The effectiveness of the proposed strategies has been extensively validated through numerous experiments.