Abstract:

The computational complexity of Bayesian Networks(BN) increases with the increase of the number of nodes, and the optimal structure of BN is still a NP(Non-deterministic Polynomial Time)-hard problem. To optimize the BN structure and improve the computing power of the complex BN structure, the BN structure is optimized through the hybrid learning method of constraints and scores. In the constrained learning, PC (Peter-Clark) algorithm is used to generate the initial network structure to improve the initial score of the network. Score-based learning uses the sparrow search algorithm to find the optimal structure of BN to enhance its scoring search ability in BN. The sparrow search algorithm and PC algorithm are applied to BN to optimize its structure, and the standard BN is used to conduct experiments, which proves the feasibility and effectiveness of the proposed algorithm in BN structure learning. Experiments on networks with different complexities show that the proposed method obtains better BIC scores than other algorithms, and in the test of 2 000 samples on the ASIA network, the error from the standard score is only 0.2.

Key words: Bayesian network, structure learning, BIC score, prior structure, sparrow search algorithm, PC algorithm, constraint-based learning, score-based learning