Partitioned hybrid learning of Bayesian network structures

Abstract We develop a novel hybrid method for Bayesian network structure learning called partitioned greedy search (pHGS), composed of three distinct yet compatible new algorithms: Partitioned PC (pPC) accelerates skeleton via divide-and-conquer strategy, p -value adjacency thresholding (PATH) effectively accomplishes parameter tuning with single execution, and initialization (HGI) maximally utilizes constraint-based information to obtain high-scoring well-performing initial graph search. establish consistency our algorithms in the large-sample limit, empirically validate methods individually collectively through extensive numerical comparisons. The combined merits pPC PATH achieve significant computational reductions compared algorithm without sacrificing accuracy estimated structures, generally applicable HGI strategy reliably improves estimation structural popular negligible additional expense. Our empirical results demonstrate competitive performance pHGS against many state-of-the-art algorithms.