Efficient Discovery of Spouses for Causal Feature Selection

The identification of Markov Blankets (MB) plays a pivotal role in the domains of structural learning and the optimization of feature selection within Bayesian Networks (BN). Although numerous algorithms for efficient or effective MB discovery have been introduced, a persistent challenge remains the reconciliation of computational accuracy with speed. This paper introduces a novel MB discovery algorithm, Efficient Discovery of Spouses for Markov Blankets discovery (ESMB), which enhances spouses identification efficiency by narrowing the candidate set's search range during the PC discovery algorithm phase, and use the separation set to further narrow the spouses search space based on the spouse sufficiency condition before using the AND rule. We evaluate the ESMB algorithm against leading MB discovery algorithms across a suite of standard Bayesian network datasets. Our findings demonstrate that ESMB achieves a balance between execution efficiency and accuracy, rivaling the fastest MB discovery algorithms in speed while approaching the accuracy of the most emblematic MB discovery algorithms. This balance is particularly pronounced in the context of large Bayesian Networks.