An Evolutionary Multitasking Multiobjective Algorithm for High-Dimensional Feature Selection: Utilizing Implicit and Explicit Information

Evolutionary computation (EC) has been successful in feature selection. However, when tackling high-dimensional feature selection (HDFS) problem, EC-based algorithms face the challenge known as the “curse of dimensionality”. Recently, many EC-based HDFS algorithms have been developed and achieve promising results. Despite that, they often fail to share information and optimize collaboratively, which may make them fall into local optima. To this end, we propose a multitasking evolutionary algorithm that utilizing the sharing of implicit and explicit information between tasks to solve the HDFS problem. Firstly, the original FS task is modeled as the multi-objective HDFS problem, which is viewed as the main task and aims to explore the entire search space. Secondly, multiple auxiliary tasks with low dimensions are constructed to solve the “curse of dimensionality”, where potential feature subsets can be searched quickly. To fully leverage the advantages of both the main FS task and the auxiliary FS tasks, a knowledge transfer strategy is developed. By sharing implicit and explicit information between tasks, high-quality feature subsets can be obtained. Experimental results on 12 high-dimensional datasets show that the proposed algorithm outperforms the state-of-the-art.