Overcoming Data Distribution Shift in AIoT Mobile Sensing

Mobile devices (e.g., smartphones, wearables, and drones) are increasingly equipped with various sensing modules, including IMU modules, magnetometers, and wireless sensors. The field of mobile sensing studies how to utilize the sensing data to infer the physical context of mobile devices (e.g., device locations and user activities), which is fundamental to sensing services spanning mobile health, location-based services, and the low-altitude economy. While the recent integration of AI technologies has advanced the data processing of mobile sensing (a paradigm known as AIoT), naively applying AI in mobile scenarios often leads to unexpected failures. This is primarily due to distribution shifts of sensing data in between model training and real-world deployment, owing to factors like user heterogeneity and environmental variations. Through strategies of adaptation and generalization learning, this thesis studies deployable approaches to overcome real-world data distribution shifts in several tasks within AIoT mobile sensing. First, we propose an adaptation approach for regression tasks by introducing TASFAR, which is a generic algorithm for source-free unsupervised domain adaptation that is both agnostic to distribution shifts and storage-efficient for mobile deployment. Second, we study generalization learning for two classification tasks by proposing ELESON and PRID. ELESON uses smartphone to reliably classifying user’s conveyor states by addressing signal corruption due to user behaviors. PRID achieves IMU-assisted BLE proximity detection and cuts the spurious correlation owing to device carriage. Finally, we extend generalization ability to sequential decision-making by proposing LoRaCompass, which is a robust reinforcement learning model that guides a LoRa sensor to efficiently search for an IoT tag across diverse, unknown environments.