KGAN-Based Semisupervised Domain Adapted Human Activity Recognition

Abstract

Human activity recognition (HAR) has become a popular stream in computing and wearable technologies, opening doors to multiple applications including healthcare and sports analysis. The primary challenge in HAR is developing models that generalize effectively across multiple domains due to variations in sensor qualities and individual differences. In addition, labeling the voluminous sensed data is a major challenge for classifying the activities. Therefore, domain adaptation approaches are being extensively investigated to help address these issues, transferring knowledge from a labeled source domain to a target domain. In this letter, a novel Kolmogorov–Arnold representation-based generative adversarial network (GAN), abbreviated as KGAN is designed for recognizing activities in a semi-supervised domain-adapted environment with enhanced robustness. The proposed KGAN-based HAR framework creates synthetic data for solving the scarcity of labeled data in the target domain and enables better handling of multidimensional data through structured function approximation. In addition, the combined use of kernel mean matching (KMM) and maximum mean and covariance discrepancy (MMCD) methods boosts the domain adapted framework by negating the weaknesses of each other. HAR experiments carried out on two publicly available datasets, namely, UCI-HAR and HHAR, exhibit enhanced performance of the proposed KGAN-based framework in target domain over few state-of-the-art models. The proposed framework significantly improves the overall accuracy of HAR across various domains by utilizing the strengths of GAN, KMM, and MMCD.