This study systematically evaluated four Ant Lion Optimizer (ALO) variants for high-dimensional feature selection, aiming to identify how each approach balances exploration and exploitation. The research addressed a gap in the current meta-heuristic literature by examining the performance of Standard ALO, Adaptive Parameter ALO, Chaotic Map ALO, Opposition-Based ALO, and ALO with Local Search across diverse real-world datasets. The methodology involved uniform initialization, binary conversion, and iterative improvement phases for each algorithm variant, with primary evaluation metrics including classification accuracy, fitness convergence, feature reduction, and execution time. The results indicated that incorporating adaptive parameter tuning, chaotic sequences, opposition-based strategies, or local search mechanisms significantly improved feature selection efficiency, defined as the algorithm’s ability to identify a compact feature subset that enhances classification accuracy while minimizing the number of selected features. These improvements stemmed from the fact that techniques such as adaptive parameters and chaotic sequences introduce new forms of exploration and exploitation, enabling the optimizer to converge on a smaller, more relevant subset of features in fewer iterations. In particular, local search refinements consistently demonstrated superior feature reduction without compromising accuracy, while chaotic initialization and opposition-based learning showed significant potential in enhancing algorithmic diversity and convergence speed. To validate each ALO variant, we tested nine heterogeneous real-world datasets spanning domains such as biology, healthcare, text classification, and image analysis, with feature counts ranging from a few hundred to over a thousand. This dataset diversity highlights the algorithms’ robustness in varied high-dimensional scenarios. The findings suggest that selecting or combining specific algorithmic modifications—such as local search or adaptive parameters—can enhance performance across different data dimensionalities. Nonetheless, constraints such as limited hyperparameter variations and the absence of parallel implementations present opportunities for future research. Overall, this systematic comparison contributes to a deeper understanding of how meta-heuristic modifications can be leveraged to manage the complexity inherent in high-dimensional feature selection tasks.