Adapter-augmented Time Series Reconstruction for Source-free Domain Adaptation: A Black-box Method

Hankang Sun, Guiming Li, Su Yang, Baoqi Li

Keywords: Time Series Classification, Transfer Learning, Domain Adaptation

TL;DR: A Black-box Method for Source-free Domain Adaptation in Time Series Classification

Abstract: Domain adaptation for time series classification is challenging due to the highly dynamic nature. This study addresses the most difficult subtask where both target labels and source data are inaccessible during adaptation, namely, source-free domain adaptation (SFDA). Although several promising approaches have been proposed, the problem remains underexplored. One issue is that most existing time series SFDA methods are tightly coupled with the architecture of the classification backbone, based on fine-tuning the backbone encoder to align with the source domain. In contrast, our method performs adaptation directly on the time series data rather than on latent features, treating the backbone classification network as a black box. This design significantly enhances the method’s generality and applicability across different architectures. Specifically, we propose a coarse-to-fine adaptation framework: First, a source-pretrained reconstructor generates a base anchor that reflects domain-shared patterns. Second, a lightweight adapter is trained to further reduce the domain shift by jointly reducing the uncertainty of classification and the reconstructive error. Here, the adaptation is performed by updating only the adapter, while the full classification backbone remains frozen, allowing parameter-efficient fine-tuning based on learned priori from pre-training. Extensive experiments validate the state-of-the-art (SOTA) performance of the proposed method. Our codes are available at https://anonymous.4open.science/r/ATSR-SFDA-52EB/

Primary Area: learning on time series and dynamical systems

Submission Number: 7417