Prototype Guided Adversarial Domain Adaptation for Time Series Classification

Unsupervised Domain Adaptation (UDA) is a powerful solution to domain shift problems by transferring knowledge from labeled source domain to unlabeled target domain. Although UDA has gained much attention in visual applications, its exploration in time series has been relatively limited. In the paper, we propose a Prototype Guided Adversarial Domain Adaptation (PGADA) method to achieve UDA for time series classification. The method is designed based on a Teacher-Student architecture and uses an attention-based meta-network to transfer knowledge from teacher to student. Afterwards, a bi-classifier adversarial strategy is then used to align the source and target domains based on the student network. Meanwhile a prototype network generates prototypes to promote category-level alignment, in which the prototypes are gradually updated by more confident target samples. A bi-classifier achieves the category decision boundary adaptively adjusting to improve the classification performance on the target domain. We conduct the experiments on the practical time series datasets and show that the superiority of the proposed method.