Elastic net – based K2 algorithm for Bayesian network structure learning
Mehryar Fallahnejad, Vahid Rezaei Tabar, Mohammad Kazemi · Statistics · 2025
The K2 algorithm is widely recognised as an effective approach for Bayesian network structure learning, but its performance is highly sensitive to the chosen node order. To mitigate this limitation, we introduce a new method for deriving candidate parents that serve as input to the K2 algorithm. Our approach begins by estimating the Markov blanket of each variable using an elastic net–penalised Markov blanket procedure. We then employ a dependency-based scoring function to identify candidate parents within each Markov blanket. These selected parents are subsequently provided to the K2 algorithm to learn the network structure. Experimental results across multiple datasets show that the proposed method reduces unnecessary edges and achieves substantially better performance than existing approaches.