Title: On-Device Personalized Human Activity Recognition via Uncertainty-Aware (Un)Supervised Few-Shot User Adaptation

Abstract: Human Activity Recognition (HAR) models built on sensor data frequently suffer performance drops when applied to new users, a issue stemming from domain shifts induced by variations in sensor positioning and individual movement styles. To address this, practical wearable HAR systems need personalization techniques that are computationally lightweight, capable of handling labeled, unlabeled, or entirely absent calibration data, and resilient when calibration samples are scarce. This paper introduces a gradient-free framework that transforms pretrained HAR classifiers into Prototypical Networks by leveraging prior prototypes. These prototypes serve to maintain zero-shot accuracy and provide regularization during the adaptation process. For scenarios with labeled calibration data, we propose a closed-form Bayesian method for estimating prototypes, a logic we subsequently adapt for unlabeled calibration contexts. Experimental results across four datasets demonstrate that supervised adaptation, using merely 3 seconds of calibration data per activity (one-shot), boosts macro-F1 scores by between +2.76 and +33.44 percentage points. Similarly, unsupervised adaptation yields improvements ranging from +0.56 to +32.13 points. Because the adaptation process relies solely on closed-form prototype updates, our approach facilitates efficient and robust on-device personalization of existing HAR classifiers.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC