Domain adaptation of time series via contrastive learning with task-specific consistency

Abstract

Unsupervised domain adaptation (UDA) for time series analysis remains challenging due to the lack of labeled data in target domains. Existing methods rely heavily on auxiliary data yet often fail to fully exploit the intrinsic task consistency between different domains. To address this limitation, we propose a novel time series UDA framework called CLTC that enhances feature transferability by capturing semantic context and reconstructing class-wise representations. Specifically, contrastive learning is first utilized to capture contextual representations that enable label transfer across domains. Dual reconstruction on samples from the same class then refines the task-specific features to improve consistency. To align the cross-domain distributions without target labels, we leverage Sinkhorn divergence which can handle non-overlapping supports. Consequently, our CLTC reduces the domain gap while retaining task-specific consistency for effective knowledge transfer. Extensive experiments on four time series benchmarks demonstrate state-of-the-art performance improvements of 0.7-3.6% over existing methods, and ablation study validates the efficacy of each component.