Unsupervised Domain Adaptation in Activity Recognition: A GAN-Based Approach

Sensor-based human activity recognition (HAR) is having a significant impact in wide range of applications smart city, home, and personal healthcare. Such deployment HAR systems often faces the annotation-scarcity challenge; that is, most techniques, especially deep learning require large number training data while annotating sensor very time- effort-consuming. Unsupervised domain adaptation has been successfully applied to tackle this challenge, where knowledge from well-annotated can be transferred new, unlabelled domain. However, these existing techniques do not perform well on highly heterogeneous domains. This article proposes shift-GAN integrate bidirectional generative adversarial networks (Bi-GAN) kernel mean matching (KMM) an innovative way learn intrinsic, robust feature transfer between two Bi-GAN consists GANs are bound by cyclic constraint, which enables more effective than classic, single GAN model. KMM powerful non-parametric technique correct covariate shift, further improves space alignment. Through series comprehensive, empirical evaluations, only achieved its superior performance over 10 state-of-the-art but also demonstrated effectiveness activity-independent, intrinsic mappings domains, robustness noise, less sensitivity data.