A Modified African Vultures Optimization Algorithm for Enhanced Feature Selection

Feature selection is a reliable technique for reducing redundant, noisy, or inaccurate features in raw input datasets without compromising classifier accuracy. Integrating metaheuristic search algorithms (MSAs) into feature selection enables the discovery of pertinent features, simplifying dataset representation. However, traditional MSAs, such as the African Vultures Optimization Algorithm (AVOA) encounter limitations with their initialization scheme, leading to a higher likelihood of local optima trapping when handling datasets with numerous features. To address this issue, a modified MSA called Multi-Chaotic African Vultures Optimization (MCAVOA) is proposed for more effective feature selection. Specifically, multiple chaotic maps are employed to generate the initial population of MCAVOA, enhancing its robustness against premature convergence in complex optimization problems like feature selection. The performances of MCAVOA are evaluated using twelve datasets from the UCI Machine Learning Repository and compared with the original AVOA, as well as four other MSAs. Extensive simulation studies reveal that MCAVOA achieves the highest mean accuracy and selects the fewest features in majority of employed datasets.