Generative User Adaptation for Sensor-Based Human Activity Recognition

Sensor-based Human Activity Recognition (HAR) using inertial data is a mature field with wide-ranging applications in healthcare, sports analytics, and mobile computing [1,3]. Deep learning classifiers have become the standard for modeling these multivariate time series, but they frequently struggle to generalize across users due to inter-subject variability, a domain shift caused by differences in physiology, behavior, and sensor placement [1].

Recent advances in HAR have demonstrated that Prototypical Networks can effectively resolve this domain shift by capturing user-specific movement traits as continuous latent prototypes using only a few seconds of calibration data [1]. While this provides a highly efficient, gradient-free adaptation mechanism for classifiers, combining it with personalized data augmentation offers a powerful path to synthesize comprehensive, user-specific training sets. Traditional diffusion models have been used to generate synthetic sensor data [3], but Stochastic Interpolants and Flow Matching [2] provide a faster, mathematically robust alternative by directly modeling straight probability flow trajectories.

This thesis proposes to systematically investigate a unified framework that couples Flow Matching [2] with Prototypical Networks [1] to generate tailored data for downstream classifier training. By utilizing dynamic, batch-wise Bayesian prototype estimation during training, the generative model learns how shifts in the prototype space map to physical changes in sensor waveforms [1]. Alternatively, at inference time, test-time trajectory steering [4] can be applied using the gradient of the distance to a new user's updated prototype to guide the generation process. Despite these promising properties, the potential of combining Flow Matching, prototype-based adaptation, and test-time steering to train personalized HAR classifiers remains largely unexplored. This thesis focuses on generating tailored waveforms to completely eliminate user-level domain shift in downstream classification.

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 literature on continuous-time generative models (Flow Matching, Stochastic Interpolants) and metric-based few-shot learning (Prototypical Networks).
• Design and implement a Flow Matching pipeline for multi-channel sensor data conditioned on dynamic, batch-wise subject prototypes via Classifier-Free Guidance (CFG).
• Investigate and implement test-time steering mechanisms to guide generative trajectories toward producing unseen target-user data.
• Generate subject-specific synthetic datasets and use them to train and personalize downstream HAR classifiers.
• Evaluate downstream classification performance on standard HAR benchmarks under supervised and unsupervised few-shot calibration regimes.
• Benchmark against standard diffusion-based personalization and classical domain adaptation baselines.

• Solid understanding of deep learning concepts, particularly continuous-time generative models (Diffusion)
• Proficient programming skills in Python and PyTorch.
• Experience working with time series or sensor data is a plus but can be acquired during the thesis.
• Interest in domain adaptation, generative models, and few-shot learning is a plus.