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

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Time-series (TS) data is crucial for IoT devices but labeling is expensive.
• Unsupervised Domain Adaptation (UDA) addresses this, with Source-Free UDA (SFUDA) emerging due to privacy concerns.
• Existing SFUDA methods struggle with TS data's temporal dependencies, especially without source data or specific pretraining.

Purpose of the Study:

• To propose Temporal Source Recovery (TemSR), a novel framework for practical TS-SFUDA.
• To enable effective transfer of temporal dependencies without accessing source data or requiring source-specific designs.
• To overcome limitations of existing SFUDA methods in handling TS data characteristics.

Main Methods:

• TemSR generates a source-like domain by leveraging intrinsic TS data properties to recover temporal dependencies.
• A masking-recovery-optimization process creates a source-like distribution with restored temporal dependencies.
• Local context-aware regularization and anchor-based recovery diversity maximization refine the distribution and preserve dependencies.

Main Results:

• TemSR effectively recovers temporal dependencies and facilitates domain transfer in TS-SFUDA.
• The framework enables adaptation to target domains without source data access or specific pretraining requirements.
• Experiments across seven TS tasks show TemSR's superior performance compared to existing TS-SFUDA methods.

Conclusions:

• TemSR provides an effective and practical solution for TS-SFUDA by recovering crucial temporal dependencies.
• The proposed method overcomes the limitations of prior approaches, offering a more flexible and applicable solution.
• TemSR demonstrates significant potential for advancing UDA techniques in the context of privacy-preserving time-series analysis.