Comparative Analysis of Metaheuristic Feature Selection in SVM-Based Human Activity Recognition

Smartphone-based human activity recognition (HAR) commonly relies on large sets of engineered time- and frequency-domain variables. Although these representations capture complementary motion patterns, redundant variables may increase model cost without resolving ambiguity between similar activities. This study compared a genetic algorithm (GA), particle swarm optimization (PSO), and grey wolf optimization (GWO) as wrapper-based feature selectors for a radial basis function support vector machine (SVM) on the UCI HAR dataset. The original 7,352/2,947 training-test split and a common SVM configuration were retained. Each optimizer searched a 561-bit feature mask with a population of 10 for 10 generations or iterations, using an objective that prioritized predictive accuracy while penalizing subset size. Performance was assessed by held-out accuracy, weighted F1, selected features, confusion patterns, convergence, and total runtime. The baseline SVM achieved 0.9308 accuracy and 0.9304 weighted F1 with all 561 features. GA+SVM produced the strongest result: 0.9508 accuracy, 0.9506 weighted F1, 300 features, and 217.99 s. This reduced dimensionality by 46.52% and test errors from 204 to 145. GWO+SVM reached 0.9471 accuracy with 341 features and 295.12 s. PSO+SVM selected the smallest subset (284 features; 49.38% reduction), but its 0.9355 accuracy and 489.68 s runtime yielded a weaker trade-off. Errors remained concentrated in upstairs/downstairs and sitting/standing/laying distinctions. Under the tested search budget, GA+SVM offered the best observed balance between predictive performance, compactness, and optimization cost. Because the stochastic algorithms were not evaluated over repeated independent runs, the ranking should be interpreted as configuration-specific rather than universal.