Selection of Optimal Sub-ensembles of Classifiers through Evolutionary Game Theory


Ensemble methodology combines multiple learning schemes in order to improve the classification performance. An ensemble made of a large number of classifiers entails an increase in the computational cost, memory storage, and even a reduction in the predictive performance. Ensemble pruning has become an important task that lives up to these challenges. The thrust consists of constructing a subset that maintains or improves the accuracy of the original set of classifiers while reducing the number of members that constitute the ensemble. Inspired by evolutionary game theory, we formulate ensemble pruning as a dynamical system. Our approach simulates the evolution of an ensemble according to the replicator dynamics. The selection mechanism is defined based on strategic interactions that favor complementary members. The survivors of the evolutionary process compose the final ensemble. In order to evaluate the proposed technique, we performed comparisons with some major state-of-the-art methods such as semi-definite programming, genetic algorithm, and orientation ordering, based on 30 UCI benchmark datasets. The results demonstrate the effectiveness of our approach in terms of accuracy performance, pruning ratio, and computational cost. Index Terms—evolutionary game theory, ensemble learning, ensemble pruning, replicator dynamics