Overcoming domain shift in mobile sensing via machine learning

Abstract

Mobile sensing utilizes sensors from mobile devices (e.g., smartphones and wearables) to infer user contexts and provide appropriate services accordingly. Integrated with deep learning, which enables the understanding of multi-dimensional sensory data, mobile sensing has broad applications ranging from human activity recognition to mobile healthcare. However, we found that various factors such as different users, devices, and environments impact the performance of such applications, thus making the domain shift (i.e., distributional shift between the training domain and the target domain) a critical issue in mobile sensing. To overcome the domain shift in mobile sensing, this thesis proposes four approaches. First, we propose a meta-learning-based adaptation framework that utilizes very few data instances from a target user. Second, we propose a similar condition detector that identifies when the unseen condition has similar characteristics to seen conditions and leverages this hint to further improve the accuracy. Third, to eliminate the labeling burden and adapt to gradually changing environments, we propose a test-time adaptation scheme that adapts to a target user's domain during inference without any data collection. Lastly, to reduce the uncertainty of the performance after adaptation, we present a framework that estimates the adaptation performance in a target domain with only unlabeled target data. Our evaluation with real-world datasets indicates that the proposed methods not only outperform the baselines but also significantly reduce computational overhead, which is crucial for resource-constrained mobile devices.