Abstract

This paper proposes a bi-objective approach to address the data-driven Bayesian network structure learning problem. The objectives considered for optimization are minimum description length (MDL) and misclassification. An algorithm based on the well-known multi-objective particle swarm optimization (MOPSO), called MOPSO-BN, is used to tackle the bi-objective learning problem. Furthermore, a strategy for preference handling from the Pareto front that selects the nearest model to a reference point is proposed. Finally, this bi-objective approach is compared against a single-objective approach. Numerical results show how this multi-objective approach is highly efficient at competitive Bayesian networks with a balanced trade-off between MDL and misclassification.

IPC Classification

Keywords