Wi-Adaptor: Fine-grained Domain Adaptation in WiFi-based Activity Recognition

Human activity recognition (HAR) has attracted significant attention during recent years due to its critical role in a wide range of applications. Among existing recognition algorithms, most of them utilize domain adversarial neural networks, such as DANN [6], to achieve recognition between diverse domains. However, these methods try to fully align the feature distributions while each domain has specific characteristics, which leads to different decision boundaries and substantially degrades the recognition accuracy. In this paper, we propose a fine-grained method called Wi-Adaptor to tackle these problems. Wi-Adaptor utilizes two classifiers to match distributions of source and target samples by considering the decision boundaries. In order to detect target samples that are far from the support of the source, we train the classifiers to maximize the discrepancy between their outputs and train the feature generator to generate target features that minimize the discrepancy. Our experiments show that Wi-Adaptor outperforms other traditional domain adversarial adaptation models and show robustness as we limit the source samples. Especially in the case of reducing the source samples to a half, Wi-Adaptor achieves more than 30% accuracy gain in different domain adaptation experiments.