Abstract
Domain shift poses a fundamental challenge in time series analysis, wheremodels trained on source domain often fail dramatically when applied in targetdomain with different yet similar distributions. While current unsuperviseddomain adaptation (UDA) methods attempt to align cross-domain featuredistributions, they typically treat features as indivisible entities, ignoringtheir intrinsic compositions that governs domain adaptation. We introduceDARSD, a novel UDA framework with theoretical explainability that explicitlyrealizes UDA tasks from the perspective of representation space decomposition.Our core insight is that effective domain adaptation requires not justalignment, but principled disentanglement of transferable knowledge from mixedrepresentations. DARSD consists three synergistic components: (I) Anadversarial learnable common invariant basis that projects original featuresinto a domain-invariant subspace while preserving semantic content; (II) Aprototypical pseudo-labeling mechanism that dynamically separates targetfeatures based on confidence, hindering error accumulation; (III) A hybridcontrastive optimization strategy that simultaneously enforces featureclustering and consistency while mitigating emerging distribution gaps.Comprehensive experiments conducted on four benchmark datasets (WISDM, HAR,HHAR, and MFD) demonstrate DARSD's superiority against 12 UDA algorithms,achieving optimal performance in 35 out of 53 cross-domain scenarios.