Cross-domain activity recognition via substructural optimal transport

It is expensive and time-consuming to collect sufficient labeled data for human activity recognition (HAR). Domain adaptation a promising approach cross-domain recognition. Existing methods mainly focus on adapting representations via domain-level, class-level, or sample-level distribution matching. However, they might fail capture the fine-grained locality information in data. The domain- class-level matching are too coarse that may result under-adaptation, while be affected by noise seriously eventually cause over-adaptation. In this paper, we propose substructure-level domain (SSDA) better utilize of accurate efficient knowledge transfer. Based SSDA, an optimal transport-based implementation, Substructural Optimal Transport (SOT), HAR. We obtain substructures activities clustering seeks coupling weighted between different domains. conduct comprehensive experiments four public datasets (i.e. UCI-DSADS, UCI-HAR, USC-HAD, PAMAP2), which demonstrates SOT significantly outperforms other state-of-the-art w.r.t classification accuracy (9%+ improvement). addition, our mehtod 5x faster than traditional OT-based DA with same hyper-parameters.