WiDA-KAN: An integrated framework for joint Wi-Fi CSI-based activity recognition and indoor localization

Human activity recognition (HAR) plays an important role in healthcare and smart home applications. However, vision-based approaches often raise serious privacy concerns, while wearable sensor-based methods are limited by inconvenience and poor long-term usability. As a contact-free and privacy-preserving sensing modality, Wi-Fi channel state information (CSI) has emerged as a promising alternative for HAR. Nevertheless, existing CSI-based methods remain vulnerable to noise and interference, struggle to capture non-stationary motion dynamics, and often suffer from optimization imbalance in joint learning scenarios. To address these challenges, we propose WiDA-KAN, a unified framework for joint activity recognition and indoor localization. The framework progressively improves CSI-based sensing by refining motion-sensitive representations under noisy conditions, aggregating temporal dynamics with a hierarchical encoder and applying a wavelet-integrated Kolmogorov–Arnold Network (KAN) to the resulting motion descriptors, and stabilizing joint training with an adaptive weighting strategy. Extensive experiments on four datasets demonstrate the effectiveness of the proposed framework. Across repeated runs, WiDA-KAN obtains mean activity recognition and localization accuracies of 95.04% and 98.60% on ARIL, 96.84% and 99.98% on CSIDA, and 99.71% and 91.04% on ElderAL-CSI. On WiMANS, the two-task average ranges from 73.29% to 77.16% across nine environment–frequency settings.