Enhancing Wearable-Based Human Activity Recognition with Binary Nature-Inspired Optimization Algorithms for Feature Selection

Abstract

This research paper explores the performance of binary nature-inspired optimization algorithms as feature selection to enhance the identification of human activities using wearable technology. Utilization of nature-inspired algorithms for feature selection, as documented in scholarly literature, presents a promising opportunity to enhance machine learning and data analysis tasks, given their effectiveness in identifying relevant features, resulting in models with reduced computational complexity, improved predictive accuracy and easier interpretation. In the experiment, we conducted an evaluation of the effectiveness and efficiency of four nature-inspired binary algorithms for optimization namely Binary Particle Swarm Optimization (BPSO), Binary Grey Wolf Optimization algorithm (BGWO), Binary Differential Evolution algorithm (BDE), and Binary Salp Swarm algorithm (BSS) - in the context of human activity recognition (HAR). The outcomes of this comprehensive experimentation, conducted on two distinct human activity recognition (HAR) datasets, provide valuable insights. BPSO algorithm emerges as an adaptable and well-rounded performer, achieving a competitive balance between feature selection quality and computational efficiency in SBHAR dataset. Conversely, for the PAMAP2 dataset, BDE algorithm displays superior feature selection quality and BPSO algorithm maintains competitive performance and adaptability. In both datasets, the nature-inspired optimization algorithms have achieved remarkable feature reduction, demonstrating reductions of 48% and 50% respectively. The experiment results show how these algorithms could be used to improve methods for recognizing human activities using wearables technology, such as feature selection, parameter adjustment, and model optimization.