The exponential growth of data in recent years has introduced significant challenges in managing high-dimensional datasets, particularly in industrial contexts where efficient data handling and process innovation are critical. Feature selection, an essential step in data-driven process innovation, aims to identify the most relevant features to improve model interpretability, reduce complexity, and enhance predictive accuracy. To address the limitations of existing feature selection methods, this study introduces a novel wrapper-based feature selection framework leveraging the recently proposed Arctic Puffin Optimization (APO) algorithm. Specifically, we incorporate a specialized conversion mechanism to effectively adapt APO from continuous optimization to discrete, binary feature selection problems. Moreover, we introduce a fully parallelized implementation of APO in which both the search operators and fitness evaluations are executed concurrently using MATLAB’s Parallel Computing Toolbox. This parallel design significantly improves runtime efficiency and scalability, particularly for high-dimensional feature spaces. Extensive comparative experiments conducted against 14 state-of-the-art metaheuristic algorithms across 15 benchmark datasets reveal that the proposed APO-based method consistently achieves superior classification accuracy while selecting fewer features. These findings highlight the robustness and effectiveness of APO, validating its potential for advancing process innovation, economic productivity and smart city application in real-world machine learning scenarios.