Abstract
Feature selection (FS) is a key process in many pattern-recognition tasks, which reduces dimensionality by eliminating redundant or irrelevant features. However, for complex high-dimensional issues, traditional FS methods cannot find the ideal feature combination. To overcome this disadvantage, this paper presents a multispiral whale optimization algorithm (MSWOA) for feature selection. First, an Adaptive Multipopulation merging Strategy (AMS) is presented, which uses exponential variation and individual location information to divide the population, thus avoiding the premature aggregation of subpopulations and increasing candidate feature subsets. Second, a Double Spiral updating Strategy (DSS) is devised to break out of search stagnations by discovering new individual positions continuously. Last, to facilitate the convergence speed, a Baleen neighborhood Exploitation Strategy (BES) which mimics the behavior of whale tentacles is proposed. The presented algorithm is thoroughly compared with six state-of-the-art meta-heuristic methods and six promising WOA-based algorithms on 20 UCI datasets. Experimental results indicate that the proposed method is superior to other well-known competitors in most cases. In addition, the proposed method is utilized to perform feature selection in human fall-detection tasks, and extensive real experimental results further illustrate the superior ability of the proposed method in addressing practical problems.