A constrained multi-objective evolutionary algorithm for multi-class instance selection

Abstract

As a data processing technology, instance selection (IS) aims to select a small number of instances with the same (or even higher) classification capability. Due to its widely applications, many IS algorithms with promising performance have been suggested. Despite that, most of existing algorithms focus on designing new IS algorithms by using different optimizing techniques, and few of them consider the imbalance among different classes in multi-class IS. To address the problem, in this paper, a constrained optimization problem is firstly formulated for multi-class IS, where the “hard constraint” and the “soft constraint” are defined to model the multi-class IS problem more accurately. Then, to solve the constrained optimization problem, a multi-objective evolutionary algorithm termed as CMOEA-MIS is proposed, by which the instance subsets with high quality could be achieved. Specifically, in CMOEA-MIS, a constraint-based solution selection strategy is developed based on the dominance relationship that considers both constraint violation and the quality of solution, and is introduced to choose the individuals in the mating pool. In addition, a two-stage based mutation strategy is also suggested in CMOEA-MIS, by which the quality of the final obtained instance subsets is further improved. Experimental results on the multi-class datasets with different characteristics have demonstrated that CMOEA-MIS can obtain multi-class instance subsets with more than 50% reduction rate, and can ensure the accuracy of each class, and can be used to train the classifiers with comparable or better performance than the state-of-the-art IS algorithms.