Improvement of chaotic maps for k-NN classification using Lévy-flight whale optimization: multiple dataset comparison

Metaheuristic feature selection is frequently used to improve distance-based classification on tabular datasets, yet the empirical value of adding chaotic dynamics to already enhanced optimizers remains unclear. This paper presents a controlled multi-dataset benchmark of a binary Whale Optimization Algorithm (WOA) feature selector combined with Lévy-flight refinement and a k-nearest neighbors (k-NN) classifier. Four fixed chaotic-map pairs, namely Logistic+Cubic, Sine+Sinusoidal, Singer+Circle and Iterative+Tent, are then applied to the same selected feature subspace and evaluated against the WOA+Lévy + k-NN baseline. The experiments use eight publicly available tabular datasets covering diagnostic, voting, signal-processing, combinatorial and chemical classification problems. The evaluation uses repeated stratified cross-validation, train-fold-only min-max scaling, a fixed k = 3 classifier, and non-parametric statistical testing with Friedman, Wilcoxon signed-rank and Holm-Bonferroni procedures. The results do not support a general claim that chaotic-map combinations improve the baseline. On the contrary, the WOA+Lévy + k-NN baseline obtains the best average rank for both accuracy and macro-F1. The Friedman test indicates a statistically detectable difference among methods for macro-F1, but the rank pattern favors the baseline, and no chaotic-map pair significantly outperforms it after Holm correction. Dataset-level analysis shows that chaotic transformations can occasionally help, as in Tic-tac-toe, but they more often degrade or leave performance unchanged, especially on Breast-cancer-wisconsin, Ionosphere, Sonar, WDBC and Wine. The contribution is therefore a reproducible boundary-condition analysis rather than a claim of a universally superior optimizer. Although the benchmark includes a raw k-NN control, a standard WOA ablation, and an EvoMapX-inspired population evolution analysis, its conclusions remain limited to the evaluated post-selection integration strategy and the four representative fixed chaotic-map pairs rather than all possible chaotic optimization designs.