| Summary: | Unsupervised domain adaptation is a machine learning framework to transform information learned from one or several source domains with many annotated samples to
unlabeled target domains. A typical unsupervised domain adaptation method is typically designed base on visual data. Solutions on time series data are less explored.
Most existing methods are based either on mapping representations from one domain
to the other or learning the features invariant to the domains by statistical restrictions.
However, focusing only on the mapping or invariant features may leave the methods vulnerable to noise since they ignore much individual information. In this work, we propose
Contrastive Adversarial Domain Disentangled Network (CADDN), a novel method that
explores improving the model adaptation performance on time series data by exploiting
each domain’s specific properties. Our primary motivation is to construct a framework
that can learn while jointly disentangling the domain-invariant and the domain-specific
features in the mean time. Contrastive learning is applied in the optimization of the
domain-specific features, targeting a beneficial and stable feature extraction. Comprehensive experimental evaluations are conducted on four benchmark time series datasets to
demonstrate the superiority of the proposed method over state-of-the-art domain adaptation solutions. A further ablation study validates the hypothesis that adding contrastive based domain-specific feature extraction will largely improve the performance compared
to only focusing on domain-invariant knowledge.
|
|---|