Evolutionary Multitasking Bi-Directional Particle Swarm Optimization for High-Dimensional Feature Selection

Feature selection is an important data processing technique, aiming to reduce the redundant and irrelevant features of data. However, as the number of features increases, feature selection algorithms based on particle swarm optimization (PSO) face the challenges of low search efficiency and huge computational consumption due to the enormous search space. A recently proposed bi-directional feature fixation (BDFF) framework for PSO has shown its effectiveness in solving high-dimensional feature selection problems, but it may mislead the particles to search in the wrong direction and requires a long time to find a small feature subset. Utilizing the prior knowledge of feature selection is expected to further enhance the performance of BDFF. Therefore, this paper first designs two tasks to introduce the prior knowledge of feature selection into BDFF while retaining its global search ability. Then, the multitasking bi-directional PSO (MBDPSO) is proposed by combining BDFF and the evolutionary multitasking optimization (EMTO) technique, which can help transfer knowledge between the two tasks effectively. Experimental results on 10 public classification datasets demonstrate that the proposed MBDPSO has an excellent performance on high-dimensional feature selection problems.