Enabling large-scale Bayesian network learning by preserving intercluster directionality

Abstract

We propose a recursive clustering and order restriction (R-CORE) method for learning large-scale Bayesian networks. The proposed method considers a reduced search space for directed acyclic graph (DAG) structures in scoring-based Bayesian network learning. The candidate DAG structures are restricted by clustering variables and determining the intercluster directionality. The proposed method considers cycles on only c(max)(<< n) variables rather than on all n variables for DAG structures. The R-CORE method could be a useful tool in very large problems where only a very small amount of training data is available.