Time series data find extensive applications in finance, healthcare, and industrial monitoring domains. However, analytical models targeting such data are subject to notable constraints imposed by the rigid independent and identically distributed (IID) assumption and the high cost of data annotation. Unsupervised Domain Adaptation (UDA) offers an effective remedy for these challenges, and Contrastive Learning (CL) has been widely integrated into UDA frameworks, owing to its robust feature representation and clustering capabilities. Nonetheless, existing CL-based UDA methods suffer from two key limitations: (1) fixed data augmentation strategies result in imbalanced intensity—excessive augmentation erodes sample semantics, while insufficient augmentation induces model overfitting; (2) distribution alignment strategies neglect hard samples which are the core carriers of domain shift, causing their domain adaptation signals to be overshadowed by a large number of normal samples and thus degrading alignment accuracy. To address these drawbacks, this paper proposes a time-series UDA algorithm, termed Adaptive Contrastive Learning Domain Adaptation (ACLDA), which incorporates two key components: (1) an adaptive feature enhancement module that integrates adaptive sample augmentation and CL, enabling the model to capture high-quality transferable features; (2) sample-level adaptive weights, introduced on the basis of class-level alignment via supervised CL, to emphasize the value of hard samples. Comparative experiments on multiple time-series datasets demonstrate that our ACLDA outperforms state-of-the-art domain adaptation methods in terms of average accuracy, verifying its superiority and providing a more robust solution for cross-domain time series analysis.
Liu et al. (Sat,) studied this question.