Diffusion-Based Personalization for 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 [1,3]. Despite strong performance, deep learning models often generalize poorly across users due to inter-subject variability in movement patterns, physiology, and sensor placement.

Recent work addresses this domain shift through few-shot adaptation, representing user-specific movement characteristics as prototypes in a learned embedding space from limited calibration data [1]. These prototypes provide a compact representation of how users differ.

This thesis investigates whether Diffusion Models (Flow Matching / Stochastic Interpolants) [2] can leverage user prototypes to address inter-subject variability in HAR. The key idea is to learn how differences in the user embedding space translate into variations in sensor signals, enabling the generation of data representative of specific users.

User prototypes will be incorporated through conditional generation using Classifier-Free Guidance (CFG), while test-time trajectory steering [4] will be explored to adapt the generative process to previously unseen users from limited calibration data. The resulting synthetic data will be used to train or adapt downstream HAR classifiers, with the goal of improving cross-user generalization and enabling efficient few-shot personalization.

References

1. Burzer, Maximilian, et al. "Uncertainty-Aware (Un) Supervised Few-Shot User Adaptation for On-Device Personalized Human Activity Recognition." arXiv preprint arXiv:2606.04798 (2026).
2. Albergo, Michael, Nicholas M. Boffi, and Eric Vanden-Eijnden. "Stochastic interpolants: A unifying framework for flows and diffusions." Journal of Machine Learning Research 26.209 (2025): 1-80.
3. Hasegawa, Tatsuhito, and Shunsuke Sakai. "Personalization of Human Activity Recognition with Cross-Conditioned Diffusion Models." International Conference on Neural Information Processing. Singapore: Springer Nature Singapore, 2025.
4. Sabour, Amirmojtaba, et al. "Test-time scaling of diffusions with flow maps." arXiv preprint arXiv:2511.22688 (2025).

• Review Diffusion Models (Flow Matching / Stochastic Interpolants), prototype-based adaptation, and HAR.
• Develop a generative model for multi-channel inertial sensor data conditioned on user prototypes.
• Investigate CFG and test-time trajectory steering for user-conditioned generation.
• Use generated data to improve cross-user generalization and few-shot personalization of HAR classifiers.
• Evaluate on standard HAR benchmarks and compare against generative augmentation, personalization, and domain-adaptation baselines.

• Strong foundations in machine learning and deep learning.
• Proficiency in Python and PyTorch.
• Familiarity with Diffusion Models or deep generative modeling.
• Ability to understand research papers and independently implement and evaluate methods.
• Experience with time-series, domain adaptation, or few-shot learning is beneficial but not required.