Hyperbolic DivideMix: Leveraging Poincaré Ball Embeddings for Robust Human Activity Recognition Under Label Noise

Human activity recognition (HAR) plays a critical role in applications such as healthcare monitoring, rehabilitation assistance, and sports performance analysis. However, sensor-based HAR datasets inevitably contain noisy labels due to ambiguous activity boundaries and annotation inconsistencies. This issue is particularly prominent for inertial measurement unit (IMU) and electromyography (EMG) signals, whose semantics are not visually interpretable and often share submotions across activities. In this work, we propose a data-centric preprocessing framework to identify confusion-prone classes and detect potentially ambiguous labels prior to downstream HAR model training. We first convert raw sensor windows into feature vectors and embed them into the Poincaré ball, where a class complexity score - derived from intraclass dispersion and raw-signal energy - is used to modulate classwise embedding magnitudes. By jointly analyzing magnitude distributions and within-class directional deviations, we localize confusion-prone categories and assign a noise likelihood score to each sample using a hybrid Poincaré-DivideMix-lite estimator. The resulting scores enable practitioners to either remove suspicious segments or apply targeted relabeling and specialized learning strategies in subsequent HAR training. Experiments on four benchmark datasets (DSADS, MHEALTH, NinaPro DB1, and HHAR) demonstrate that our approach consistently improves noisy-sample detection, achieving average gains of +6.8% in Top-K (e.g., Top-10%) hit rate and +9.5% in recall over DivideMix, while outperforming ELR+ and C2D. The improvements are most pronounced for semantically overlapping activities (e.g., running and jumping), highlighting the suitability of hyperbolic geometry for modeling domain-specific label ambiguities in HAR.