KTOBS: An Approach of Bayesian Network Learning Based on K-tree Optimizing Ordering-Based Search

Qingwang Zhang, Sihang Liu, Ruihong Xu, Zemeng Yang, Jianxiao Liu

Abstract: How to construct Bayesian Networks (BN) efficiently and accurately is a research hotspot in the era of artificial intelligence. By limiting the tree-width of the network, the Bayesian network learning based on k-tree can be used to process large-scale of variables. However, this method has the problems of low accuracy, further to optimize the order of adding nodes, etc. In order to solve these problems, this work proposes a Bayesian learning method based on k-tree optimizing ordering-based search (KTOBS). Firstly, the local learning search strategy is adopted to obtain the candidate parent sets of each variable efficiently and accurately. Then it selects k + 1 nodes based on the obtained candidate parent node sets, and constructs the corresponding initial sub-network. Then the heuristic evaluation strategy is used to add subsequent nodes successively, and thus to get the initial network. Finally, it optimizes the network iteratively through switching nodes until the score of network no longer increases. The experimental results show that KTOBS can learn a network structure with higher accuracy than other k-tree algorithms in a given limited time. Availability and implementation: codes and experiment dataset are available at: http://122.205.95.139/KTOBS/ .