Designing High Reliable Ensemble Classifiers Using Heuristic Algorithms

The measure of reliability in decision of a classifier is sometimes more important than its recognition rate. Military and security applications are clear examples to show the importance of this factor. In this paper, a method is described for designing ensemble classifiers with high reliability. The proposed method is established based on multi-objective optimization techniques. After analyzing the measure of reliability in a classifier, the multi-objective optimization algorithms are utilized to optimize reliability, error rate, and ensemble size simultaneously. Multi-Objective Inclined Planes Optimization algorithm (as a novel multi-objective technique) and Multi-Objective Particle Swarm Optimization algorithm (as a conventional one) are the multi-objective heuristic methods applied in the present research. Using the proposed method, it's possible to create various and user-defined conditions, due to the ability of multi-objective optimization algorithm in the presentation of the Pareto front; conditions in which the importance of each factor can be strengthened and weakened. The experimental results confirm the requirement of considering the measure of reliability in ensemble classifiers; in this case, the reliability for each class is higher than reliability values obtained in the other case (without considering reliability as an objective function). It's worth noting that adding this measure will not destroy two other measures. Also, superiority of Multi-Objective Inclined Planes Optimization is seen in obtained results for all datasets; this algorithm leads to improvements in the average values of reliability, error rate, and ensemble size by 21.5%, 76.9% and 53.8% for test dataset in the best condition.