Time series data is increasingly prevalent in diverse sectors such as finance, IoT, and healthcare, with notable applications in neuroscience. Although neural networks exhibit proficiency in handling time series data, domain shift often impedes their effectiveness. To address this issue, we propose an innovative approach called Hierarchical Layer-wise Contrastive Learning for Unsupervised Domain Adaptation on Time-Series (TS-HCL). TS-HCL addresses three key aspects: cross-domain sample similarity, interference from noisy domain labels, and conditional distribution shifts. Firstly, commonalities are established across domains by treating domain feature representations at corresponding layers as positive pairs through domain-level contrastive learning. Secondly, Environment Label Smoothing (ELS) is introduced, encouraging themarginal discriminator to estimate soft probabilities, thereby alleviating the impact of domain label noise. Lastly, a conditional domain discriminator is designed to provide enhanced context and align conditional distributions. The proposed TS-HCLmethod exhibits performance in cross-domain scenarios, as demonstrated by its effectiveness across both public and private datasets, with particular excellence in medical applications.
