SEnsor Alignment for Multivariate Time-Series Unsupervised Domain Adaptation

Unsupervised Domain Adaptation (UDA) methods can reduce label dependency by mitigating the feature discrepancy between labeled samples in a source domain and unlabeled similar yet shifted target domain. Though achieving good performance, these are inapplicable for Multivariate Time-Series (MTS) data. MTS data collected from multiple sensors, each of which follows various distributions. However, most UDA solely focus on aligning global features but cannot consider distinct distributions sensor. To cope with such concerns, practical adaptation scenario is formulated as (MTS-UDA). In this paper, we propose SEnsor Alignment (SEA) MTS-UDA to at both local sensor levels. At level, design endo-feature alignment align their correlations across domains, whose information represents interactions sensors. Further, exo-feature enforce restrictions features. Meanwhile, also incorporates essential spatial-temporal dependencies be transferred existing methods. Therefore, model multi-branch self-attention mechanism simple effective transfer domains. Empirical results demonstrate state-of-the-art performance our proposed SEA two public datasets MTS-UDA. The code available https://github.com/Frank-Wang-oss/SEA