Bayesian network structure decomposition learning is an advanced method for building complex Bayesian network structure. The methods build an initial graph at one time, then decompose it and learn these sub-graphs, and finally combine all the subgraphs into a complete Bayesian network structure. However, the result is affected by the initial graph. The paper firstly proposes a novel structure decomposition learning method of complex Bayesian networks based on iterative learning. Firstly, we develop an iterative framework for complex Bayesian network structures decomposition learning. In each iteration, we continuously use the undirected graph contained in the learning result for new decomposition learning to optimize the learning result. Secondly, we propose three mutation operators to generate new initial graphs in iterative processes. Within our framework, mutation operators facilitate the generation of new initial graphs by strategically utilizing existing structural information in various manners, thereby enhancing the efficiency of the iterative learning process. Finally, numerous experiments have been conducted to demonstrate the learning process. Compared with advanced algorithms, our method outperforms non-iterative algorithms.
Jia et al. (Wed,) studied this question.