Bayesian Network Structure Learning Based On Graph Partitioning

Bayesian network as a graphical modeling tool provides a representation of a causal relationship between variables, and plays an important role in uncertain reasoning area. It is widely used in expert systems, artificial intelligence, machine learning and other fields. Structure learning is the organic combination between graph theory and probability theory. As the emphasis and difficulty in learning Bayesian network, it is a method to determine the causal relationship in the Bayesian network based on observation data from variables. With the advent of the era of big data, the scale of the network grows fast. The problems of traditional structure learning algorithm in efficiency and accuracy increasingly reveal.This paper starts from the traditional Bayesian network structure learning framework. Complex network partitioning method is added to the structure of the learning process, and the framework of learning is improved; we proposed the chaotic mix improved particle swarm optimization algorithm(CIPSO) which effectively solves the problem of prematurity. The main work:First, the methods of Bayesian network structure learning are reviewed. We divide those methods into three categories: 1, learning method based on statistical analysis; 2, learning method based on structural score search; 3, a hybrid Bayesian network structure learning method. The problems of structure learning are pointed out and at the same time the development direction is given.Then, we improved the Bayesian network structure learning framework. For large Bayesian networks, the structure space has increased exponentially with the increasing of nodes. In this case, the efficiency of traditional two-stage framework significantly reduced. This paper combines complex network partitioning method into the framework of learning and divides the whole network into multiple sub-networks. The structure learning of sub-networks works independently. Thus the learning framework is improved and the efficiency of learning is greatly increased.Finally, a CIPSO algorithm is proposed. For the simplicity of model and encoding, PSO algorithm is widely used in search scoring process. But the traditional PSO algorithm easily gets into prematurity. In this paper, we divide particle swarm into several populations. The velocity of population is considered into the update process. The chaos strategy is taken to map the particles which fall into local optima. Those strategies effectively inhibited the particle swarm into the precocious state. The results show that this method improves the accuracy of structure learning.