Temporal Source Recovery for Time-Series Source-Free Unsupervised Domain Adaptation

TL;DR

TemSR addresses the challenge of transferring temporal dependencies in time-series data under source-free unsupervised domain adaptation by generating a source-like distribution through a masking-and-recovery process guided by entropy minimization from a fixed source encoder. The framework enhances recovery with local context-aware regularization and a diversity-enhancing anchor-based module backed by an anchor bank, followed by standard distribution alignment to transfer temporal dependencies to the target. Across HAR, SSC, MFD, HHAR, and WISDM, TemSR achieves state-of-the-art performance in many cross-domain scenarios, often outperforming methods that rely on source data or source-specific pretraining, while maintaining practical compute requirements. The work also analyzes distribution discrepancies during adaptation and demonstrates TemSR’s ability to narrow the source-target gap without source access, highlighting its practical impact for privacy-preserving TS-SFUDA applications.

Abstract

Time-Series (TS) data has grown in importance with the rise of Internet of Things devices like sensors, but its labeling remains costly and complex. While Unsupervised Domain Adaptation (UDAs) offers an effective solution, growing data privacy concerns have led to the development of Source-Free UDA (SFUDAs), enabling model adaptation to target domains without accessing source data. Despite their potential, applying existing SFUDAs to TS data is challenging due to the difficulty of transferring temporal dependencies, an essential characteristic of TS data, particularly in the absence of source samples. Although prior works attempt to address this by specific source pretraining designs, such requirements are often impractical, as source data owners cannot be expected to adhere to particular pretraining schemes. To address this, we propose Temporal Source Recovery (TemSR), a framework that leverages the intrinsic properties of TS data to generate a source-like domain and recover source temporal dependencies. With this domain, TemSR enables dependency transfer to the target domain without accessing source data or relying on source-specific designs, thereby facilitating effective and practical TS-SFUDA. TemSR features a masking recovery optimization process to generate a source-like distribution with restored temporal dependencies. This distribution is further refined through local context-aware regularization to preserve local dependencies, and anchor-based recovery diversity maximization to promote distributional diversity. Together, these components enable effective temporal dependency recovery and facilitate transfer across domains using standard UDA techniques. Extensive experiments across multiple TS tasks demonstrate the effectiveness of TemSR, which even surpasses existing TS-SFUDA methods that require source-specific designs.

Figures (10)

• Figure 1: (a) The source model is trained via entropy minimization, which reduces prediction uncertainty and encourages deterministic outputs for domains with source characteristics; (b) Due to domain gaps, the fixed pretrained model yields high-entropy (uncertain) outputs on target data; (c) Masking introduces diversity into the target domain, implicitly providing candidate domains. To regain deterministic outputs, the input to the pretrained source model must exhibit source characteristics. Thus, entropy minimization inversely pushes the recovery model to generate source-like distributions.
• Figure 2: Overall TemSR. An encoder pretrained on a source domain is transferred to a target domain for adaptation without source data, using source-like and target branches. In the source-like branch, masked target samples are recovered by a recovery model. With the fixed source encoder, their entropy is computed via a local context-aware regularization loss $\mathcal{L}_{LCA}$ and minimized for optimization to generate a source-like distribution with restored temporal dependencies. Meanwhile, an Anchor-based Recovery Diversity Maximization loss $\mathcal{L}_{ARDM}$ enhances the diversity of the generated distribution for effective recovery. Finally, source-like and target distributions are aligned with an alignment loss $\mathcal{L}_{Align}$, enabling the transfer of temporal dependencies.
• Figure 3: (a) Source and target distributions are distinct but related. (b) Source-like distribution, when initialized from the target distribution, can more easily be optimized to resemble source distribution.
• Figure 4: (a) Recovery diversity maximization causes recovered samples to deviate in unintended directions without proper constraints. (b) Anchors act as reference points, balancing diversity with fidelity to the source domain.
• Figure 5: Analysis for $\lambda_{LCA}$.

Theorems & Definitions (2)

• Theorem 1
• Theorem 2