Quantifying User-Level Domain Shift in Human Activity Recognition

Sensor-based Human Activity Recognition (HAR) uses inertial sensor data to recognize human activities, with applications in healthcare, sports, and mobile computing. Despite strong benchmark performance, HAR models often generalize poorly to unseen users due to inter-subject variability in movement patterns, physiology, and sensor placement [1]. While this user-level domain shift is well known, its magnitude is difficult to quantify independently of a trained HAR model. A model-free measure of domain shift could help characterize datasets, identify challenging users, and estimate generalization difficulty before deploying a model to a new user.

This thesis investigates statistical methods for quantifying user-level domain shift directly from inertial sensor data. Sensor windows will be represented using interpretable time-, frequency-, and time-frequency features, optionally combined with classical dimensionality reduction such as Principal Component Analysis (PCA) [2]. User-specific structure will be analyzed using measures of intra-user compactness and inter-user separation, such as the Silhouette Coefficient [3], as well as distributional distances such as Maximum Mean Discrepancy (MMD) [4] and Sliced Wasserstein Distance [5]. Particular attention will be given to activity-conditional measures that distinguish differences in how users perform individual activities from differences in their overall activity distributions. The resulting metrics require no learned representation or HAR model and can therefore be applied directly to new datasets and users.

The central research question is whether these model-free domain-shift measures reflect the actual difficulty of generalizing HAR models to unseen users. The proposed metrics will therefore be benchmarked against the performance gap between unseen data from users represented during training and data from entirely unseen users, using multiple HAR models and datasets. The overall goal is to develop an interpretable, model-independent domain-shift metric that correlates with downstream HAR performance degradation and can serve as an indicator of expected cross-user generalization.

References

[1] Morales, J., et al. “In Shift and In Variance: Assessing the Robustness of HAR Deep Learning Models Against Variability.” Sensors, 25(2), 430, 2025.

[2] Jolliffe, I. T., & Cadima, J. “Principal Component Analysis: A Review and Recent Developments.” Philosophical Transactions of the Royal Society A, 374(2065), 20150202, 2016.

[3] Rousseeuw, P. J. “Silhouettes: A Graphical Aid to the Interpretation and Validation of Cluster Analysis.” Journal of Computational and Applied Mathematics, 20, 53–65, 1987.

[4] Gretton, A., et al. “A Kernel Two-Sample Test.” Journal of Machine Learning Research, 13, 723–773, 2012.

[5] Nietert, S., Sadhu, R., Goldfeld, Z., & Kato, K. “Statistical, Robustness, and Computational Guarantees for Sliced Wasserstein Distances.” 2022.

• Review inter-subject variability in HAR and model-free methods for quantifying domain shift.
• Extract and analyze time-, frequency-, and time-frequency features from inertial sensor data.
• Investigate dimensionality reduction, clustering/separability measures, and statistical distances for quantifying inter- and intra-user variability.
• Develop and analyze activity-conditional and user-level domain-shift metrics across HAR datasets.
• Benchmark the metrics against the seen-to-unseen-user performance gap of HAR models and evaluate their correlation with generalization performance.

• Basic knowledge of data science, statistics, or signal processing.
• Good programming skills in Python.
• Familiarity with data analysis, visualization, or clustering methods.
• Interest in working with time-series and sensor data.