Source-Free Time-Series Domain Adaptation With Prior Evaluation of Model Salience.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Source-free domain adaptation (SFDA) is crucial for adapting models to new data without accessing original sensitive data.
• Existing SFDA methods often neglect time-series specific features like temporal dependencies.
• Current SFDA fine-tuning is limited to output-level adaptation, lacking interpretability and risking negative transfer.

Purpose of the Study:

• To develop an interpretable SFDA framework for time-series data.
• To address the limitations of black-box posterior adaptation in SFDA.
• To improve model adaptation by incorporating semantic interpretability.

Main Methods:

• Introduced 'model salience' as a quantifiable measure of semantic interpretability.
• Developed the PrEPoA (Prior Evaluation of model salience with Posterior Adaptation) framework.
• Incorporated a Key Pattern Reconstruction (KPR) module for salience quantification and an interpattern triplet loss for calibration.
• Utilized Robust Prototype Clustering (RPC) for generating reliable pseudo-labels in the adaptation stage.

Main Results:

• PrEPoA demonstrated superior performance across multiple time-series datasets (WISDM, HAR, HHAR, MFD, SSC).
• Outperformed nine unsupervised domain adaptation (UDA) and seven SFDA methods.
• Validated PrEPoA's effectiveness as a plug-and-play module for existing SFDA methods.

Conclusions:

• The proposed PrEPoA framework offers a significant advancement in interpretable SFDA for time-series data.
• Integrating prior evaluation of model salience with posterior adaptation enhances model robustness and performance.
• PrEPoA provides a flexible and effective solution for secure and interpretable domain adaptation.