A hyper-heuristic evolutionary algorithm for learning Bayesian Network Classifiers

Abstract

Hyper-heuristic evolutionary algorithms (HHEA) are successful methods for selecting and building new heuristics or algorithms to solve optimization or machine learning problems. They were conceived to help answer questions such as given a new classification dataset, which of the solutions already proposed in the literature is the most appropriate to solve this new problem? In this direction, we propose a HHEA to automatically build Bayesian Network Classifier (BNC) tailored to a specific dataset. BNCs are powerful classification models that can deal with missing data, uncertainty and generate interpretable models. The method receives an input a set of components already present in current BNC algorithms and a specific dataset. The HHEA then searches for the best combination of components according to the input dataset. Results show the customized algorithms generated obtain results of F-measure equivalent or better than other state of the art BNC algorithms.