A New Evolutionary Multitasking Algorithm for High-Dimensional Feature Selection

Feature selection (FS) is an important dimension reduction technique in practical applications, which has been widely studied during the past decades. Although a large number of FS algorithms have been proposed and shown the promising performance, most of them face with the challenge of “curse of dimensionality”. To this end, inspired by evolutionary multitasking (EMT), in this paper, a VariAble MultiTasking-based Multi-Objective Evolutionary Algorithm, named VAMT-MOEA, is proposed for high-dimensional FS. For the existing EMT-based FS algorithms, they adopt the single or fixed assisted tasks to solve the high-dimensional FS problem (namely the original task). Once they trap into the local optima, it is difficult for them to provide valuable knowledge. Different from them, the proposed algorithm employs the variable multitasking scheme to achieve the feature subsets with high quality. The assisted task is adjusted with the changes of the weights in the evolution, where the weight measures the importance of each feature. Specifically, a variable-weight adjustment strategy is proposed to adjust the assisted task, aiming to overcome the loss of diversity during the evolution. Additionally, a novel knowledge transfer strategy is suggested, where the best and the worst solutions are used to implement the positive knowledge transfer between the assisted task and the original task. Finally, an initialization strategy is designed to generate an initial assisted task with competitiveness. The experimental results on 10 high-dimensional datasets, whose dimension ranges from 2,000 to 19,993, demonstrate the superiority of the proposed VAMT-MOEA in terms of the classification error, the number of selected features and the running time. To be specific, compared with five EA-based FS algorithms, the proposed VAMT-MOEA can achieve the feature subsets with lower classification error and smaller number of features. Moreover, the running time of different algorithms also reveal that our VAMT-MOEA uses the minimum time on all 10 high-dimensional datasets.