Assessing Distribution Shift in Human Activity Recognition for Domain Generalization

Research Question and Approach

This paper investigates the challenge of domain generalization in Human Activity Recognition (HAR). In real-world deployments, HAR models often fail because the training data (e.g., collected in a lab) does not match the distribution of the test data (e.g., different devices, sensor placements, or user behaviors). The authors systematically evaluate four types of distribution shifts—device type, sensor placement, sampling rate, and user behavior—to understand why models struggle to generalize. They introduce a new, uniform benchmark platform and evaluate 28 different domain generalization algorithms to measure their effectiveness against these shifts.

Key Findings

The authors find that all studied distribution shifts are primarily defined by 'diversity shifts,' meaning that different domains contain unique, non-shared features that make cross-domain generalization inherently difficult. When testing 28 state-of-the-art domain generalization methods, the results were sobering: these complex algorithms provided only marginal performance gains over a simple empirical risk minimization (ERM) baseline. This suggests that current techniques are not yet robust enough to handle the heterogeneity inherent in real-world sensor data.

Most HAR research assumes that training and testing data share similar distributions, an assumption that rarely holds in practice. By quantifying the specific impact of sensor-related and user-related shifts, this work highlights a significant gap in the field. The provided open-source benchmark platform and datasets serve as a critical resource for future research, encouraging the development of more robust models that can function reliably across diverse, unseen real-world conditions.