High-dimensional feature selection problems (HFSPs) are becoming more prevalent but complex. Currently used feature selection (FS) techniques for problems like biological and medical data often struggle due to the complexity of the problems they face. To adeptly address this kind of problem, a new binary form of the chameleon swarm algorithm (BCSA) called binary heterogeneous cognitive learning CSA (BHCLCSA) is proposed. In BHCLCSA, chameleons are specifically divided during optimization into topmost chameleons ( \(\text {TC}\) ) and undermost chameleons ( \(\text {UC}\) ) as per their fitness. These two types of chameleons are addressed variously by creating an elite cognitive learning (CL) mechanism to evolve the undermost chameleons and a predominant CL mechanism to mature the topmost ones. With the collaboration between these two learning strategies, BHCLCSA is expected to evolve chameleons capable of efficiently exploring the search space and exploiting the discovered optimal regions to obtain the best solutions for HFSPs. Further, this work develops adaptive and dynamic swarm partition mechanisms to adaptively segregate chameleons into two categories. HCLCSA is used in FS domain to confront premature convergence, locate the best subset of features between classes, and improve the CSA’s global and local search capabilities. The BHCLCSA-based FS method was evaluated, using the k-nearest neighbor (k-NN) classifier, on 20 HFSPs collected from the UCI repository. The results showed that BHCLCSA performed substantially better than several other widely recognized FS methods. Notably, it excelled many others in 10 datasets concerning classification accuracy, 13 datasets concerning F1-score, 7 datasets concerning the number of selected features, and 4 datasets concerning fitness values out of the 20 datasets considered. Simply put, the results on 5 and 8 datasets showed that BHCLCSA performed at performance levels exceeding 90% for F1-score and classification accuracy metrics, respectively.
