An efficient local-to-global Bayesian network skeleton learning algorithm

doi:10.1016/j.eswa.2025.127593

An efficient local-to-global Bayesian network skeleton learning algorithm

2025-07-01

0

PRE

AI

DOI：10.1016/j.eswa.2025.127593

Original

Original request for help

Save

Abstract

En 中文

Bayesian network (BN) skeleton learning is of great importance for constructing and applying BNs and contributes to causal discovery, feature selection, and many machine learning subareas. Extensive BN skeleton learning algorithms have been proposed, and most of them learn BN skeletons by performing conditional independence (CI) tests. However, those CI test-based algorithms require a large number of CI tests, including many high-order ones that cause a heavy computational burden and are prone to unreliability, which heavily impacts the efficiency and accuracy of BN skeleton learning. To tackle this issue, we propose a novel local-to-global BN skeleton learning (LGSL) algorithm, which avoids performing high-order CI tests during the learning process. LGSL learns a BN skeleton by piecing together the results obtained by LSLearner, a local BN skeleton learning algorithm proposed in this paper. LSLearner can learn the local BN skeleton of a given target node only by measuring the strength of dependency between nodes (i.e., zero-order CI tests) and comparing the magnitudes of some of the strength results. We theoretically prove the effectiveness of the proposed algorithms and conduct extensive experiments to compare the performance of LGSL with that of prototypical or state-ofthe-art BN skeleton learning algorithms on synthetic and real-world BNs. The experimental results demonstrate that LGSL dramatically improves the efficiency of BN skeleton learning without compromising accuracy.

Keywords:

Bayesian network

Skeleton learning

Structure learning

Causal discovery

An efficient local-to-global Bayesian network skeleton learning algorithm

Abstract

Journal

Organization