Domain adaptation for robust WiFi sensing: empirical analysis of domain shift and class imbalance in WiFi CSI-based human activity recognition

Abstract

WiFi-based human activity recognition (HAR) has emerged as a promising device-free sensing technology for smart homes, healthcare monitoring, and ambient assisted living. However, these models suffer from performance degradation when applied to new environments because of domain shift caused by changes in room layout, multipath propagation, and line-of-sight (LOS) versus non-line-of-sight (NLOS) conditions. This thesis addresses the robustness gap by extending Dual Adversarial Network for Human Activity Recognition (DA-HAR) with two components: a Conditional Domain Adversarial Network (CDAN) for aligning joint feature and prediction distributions and a class-weighted learning strategy to mitigate the effects of class imbalance. The proposed framework is evaluated on a public WiFi CSI dataset with 12 fine-grained activities collected in three environments (corridor, office, classroom) under LOS and NLOS conditions. Comprehensive experiments are conducted using three transfer scenarios (E1,E2→E3 / E1,E3→E2 / E2,E3→E1). The results indicate that the enhanced DA-HAR consistently outperforms both the original DA-HAR and the source-only baseline, achieving up to 7.1% absolute and 17.2% relative accuracy improvements. The confusion matrix shows that DA-HAR + CDAN reduces misclassification among fall-related and locomotion activities, indicating that it preserves class structure under domain shift better. To isolate the effect of class imbalance, an experiment is performed on 20 random six-activity subsets with and without the class-weighted loss. Class weighting improves accuracy from 0.6806 to 0.7112 and increases CDAN improvement from 5.3% to 8.7%, which means that class imbalance negatively impacts both the source model and the adaptation process. However, performance for fall activities depends on the scenario, and bending is still difficult to recognize, indicating remaining challenges. This work provides empirical evidence that robust domain adaptation for WiFi sensing requires both conditional alignment of class-specific distributions and explicit treatment of class imbalance. The proposed method offers a principled step toward reliable activity recognition in real-world environments.