Maximilian Burzer

HAR models often perform well on users represented during training but degrade substantially when applied to unseen users due to differences in individual movement patterns. However, it remains difficult to quantify how large this user-level domain shift is without relying on a trained model. This bachelor’s thesis investigates model-free statistical measures based on time- and frequency-domain sensor features, dimensionality reduction, and clustering/separability analysis. The goal is to develop an interpretable domain-shift metric that correlates with the performance drop of HAR models on unseen users and can help estimate generalization difficulty directly from the data.

HAR models often perform well on users represented during training but generalize poorly to unseen users due to differences in individual movement patterns. Collecting sufficient labeled data to retrain or personalize models for every new user is impractical. This master’s thesis investigates Diffusion Models (Flow Matching / Stochastic Interpolants) as a way to generate personalized sensor data from only limited user information. User prototypes will be used to condition and steer the generative process toward specific or previously unseen users. The resulting synthetic data will be used to improve cross-user generalization and enable few-shot personalization of HAR classifiers. Candidates should have strong deep learning foundations, Python and PyTorch skills, and an interest in generative modeling and domain adaptation.