Comparative Analysis of Support Vector Machine Variants for Human Activity Recognition Using Wearable Sensor Data

Human Activity Recognition (HAR) using wearable sensor data is widely applied in healthcare, smart environments, and mobile systems. Support Vector Machines (SVMs) are commonly used for HAR due to their strong generalization ability. However, traditional approaches often rely on single kernels and may not fully capture complex motion patterns. This research presents a comparative analysis of seven SVM-based methods, including Multiclass SVMs, Radial Basis Function (RBF) SVMs, Kernel Engineering SVMs, Multiple Kernel Learning (MKL) SVMs, Class-Weighted SVMs, Least Squares SVM approximation, and Online Incremental SVMs. A unified experimental framework with consistent preprocessing and hyperparameter tuning using Grid Search with cross-validation is employed to ensure fair evaluation. Results show that kernel-based methods outperform linear and approximate models. The MKL SVM achieves the highest accuracy, slightly surpassing the RBF baseline, by combining multiple kernels to capture diverse data characteristics. Kernel Engineering SVM also improves performance, while Fuzzy and LS-SVM provide competitive results with enhanced robustness. In contrast, Multiclass and Online SVM exhibit lower accuracy. Thes results demonstrate that improving feature representation through advanced kernel design is key to enhancing HAR performance.