Selective Cross-Domain Consistency Regularization for Time Series Domain Generalization

Domain generalization aims to learn models robust to domain shift, with limited source domains at training and without any access to target domain samples except at test time. Current domain alignment methods seek to extract features invariant across all domains, but do not consider inter-domain relationships. In this paper, we propose a novel representation learning methodology for time series classification that selectively enforces prediction consistency between source domains estimated to be closely-related. Specifically, we view a domain shift as a form of data transformation that preserves labels but not necessarily class relationships, and we regularize the predicted class relationships to be shared only by closely-related domains instead of all domains to prevent negative transfer. We conduct comprehensive experiments on two public real-world datasets. The proposed method significantly improves over the baseline and achieves better or competitive performance in comparison with state-of-the-art methods.