Decomposition-based Bayesian network structure learning algorithm using local topology information

Abstract

Hybrid learning algorithms which integrate the merits of the constraint-based methods and the search-and-score methods are used to cope with Bayesian network (BN) structure estimation problem. However, such simple and crude synthesis techniques always consider the global topology information during the learning process and attempt to directly search for the optimal network structure in the enormous solution space for large-scale BNs, resulting in prohibitive computational cost as well as low learning accuracy. Therefore, we propose a novel hybrid structure learning algorithm based on the idea of model decomposition, which takes into account the knowledge of local neighborhood structures. The proposed method works in four stages. We first draft an undirected independence graph by using an efficient Markov blanket discovery approach, and then split the entire network into a series of subgraphs. After learning the small BNs from the observed data, the resultant topology can be obtained by combining these small BNs. Experiments on different benchmark BNs and the varying data sets demonstrate that the proposed algorithm generally gains the better performance of structure recovery than other representative methods, especially for large-scale BNs.