<p>Feature selection presents a formidable optimization challenge due to its non-deterministic polynomial-time nature, despite its critical role in developing highly accurate and low-complexity machine learning models through the elimination of redundant and irrelevant features. The teaching-learning-based optimization (TLBO) algorithm is known for its robust global search capability but often suffers from premature convergence when tackling complex optimization problems. This paper proposes an enhanced variant, termed teaching-learning-based optimization with multiple superior learners interaction (TLBO-MSLI), designed as a sophisticated global search algorithm within a wrapper-based feature selection framework. TLBO-MSLI introduces substantial methodological advancements by refining both the teacher and learner phases to achieve an improved exploration-exploitation balance, thereby contributing to more efficient process innovation in data-driven model development. The multiple superior learners-guided teacher phase enables personalized guidance through diverse directional insights, while the multiple peer enhancement-guided learner phase promotes deeper and more effective knowledge exchange through interactions with elite peers. Extensive simulation studies on twenty benchmark datasets demonstrate that TLBO-MSLI attains the highest classification accuracy on thirteen datasets and ranks second on three additional ones. It also selects the smallest feature subsets in ten datasets and ranks second in four others, outperforming eleven state-of-the-art metaheuristic search algorithms. These results confirm the superior performance, consistency, and broad applicability of TLBO-MSLI, while highlighting its potential to enhance economic productivity through improved computational efficiency and optimized resource utilization in machine learning applications.</p>