Domain Adaptation Under Behavioral and Temporal Shifts for Natural Time Series Mobile Activity Recognition
This work addresses the challenge of adapting activity recognition models to real-world behavioral and temporal shifts for applications like health monitoring, but it is incremental as it builds on existing domain adaptation methods.
The authors tackled the problem of mobile activity recognition in natural settings by collecting a dataset with variations across participants and time, and found that unsupervised domain adaptation, enhanced with contrastive learning and weak supervision, improved performance, though no concrete numbers were provided.
Increasingly, human behavior is captured on mobile devices, leading to an increased interest in automated human activity recognition. However, existing datasets typically consist of scripted movements. Our long-term goal is to perform mobile activity recognition in natural settings. We collect a dataset to support this goal with activity categories that are relevant for downstream tasks such as health monitoring and intervention. Because of the large variations present in human behavior, we collect data from many participants across two different age groups. Because human behavior can change over time, we also collect data from participants over a month's time to capture the temporal drift. We hypothesize that mobile activity recognition can benefit from unsupervised domain adaptation algorithms. To address this need and test this hypothesis, we analyze the performance of domain adaptation across people and across time. We then enhance unsupervised domain adaptation with contrastive learning and with weak supervision when label proportions are available. The dataset is available at https://github.com/WSU-CASAS/smartwatch-data