DuLPA: Dual-Level Prototype Alignment for Unsupervised Domain Adaptation in Activity Recognition from Wearables

In wearable human activity recognition (WHAR), models often falter on unseen users due to behavioral and sensor differences.Without target labels, unsupervised domain adaptation (UDA) can help improve cross-user generalization. However, many WHAR UDA methods either pool all source users together or perform one-to-one source–target alignment, ignoring individual differences and risking negative transfer. To address this critical limitation, we propose \textbf{\textit{DuLPA}}—\underline{\textbf{Du}}al-\underline{\textbf{L}}evel \underline{\textbf{P}}rototype \underline{\textbf{A}}lignment method for unsupervised cross-user domain adaptation. First, it aligns class prototypes between each source user and the target to capture individual variation; a convex reweighting further handles class imbalance. Second, a BLUP-based fusion forms robust global class prototypes by optimally weighting domain-specific ones using estimated within- and between-domain variances. On four public datasets, \textbf{\textit{DuLPA}} outperforms several baselines, improving macro-F1 by 5.34\%.