Time-series domain adaptation via sparse associative structure alignment: Learning invariance and variance.

Area of Science:

• Machine Learning
• Time-Series Analysis
• Causal Inference

Background:

• Domain adaptation for time-series data is crucial for industrial applications like anomaly detection and forecasting but is under-explored academically.
• Existing methods often rely on covariate shift assumptions, which are frequently violated in time-series due to complex dependencies and sensitivity to time lags.
• The stability of causal structures across domains offers a promising alternative for robust domain adaptation.

Purpose of the Study:

• To develop a novel unsupervised domain adaptation method for time-series data that leverages stable associative structures.
• To enhance domain adaptation by incorporating domain-specific information alongside domain-invariant structures.
• To improve the performance and interpretability of time-series domain adaptation models.

Main Methods:

• Proposed the Sparse Associative Structure Alignment model with domain-specific information enhancement (SASA2).
• Generated segment sets to mitigate offset issues and extracted unweighted sparse associative structures using sparse attention mechanisms.
• Incorporated domain-specific information via an autoregressive module and employed unidirectional alignment for source-to-target transformation.

Main Results:

• SASA2 achieved state-of-the-art performance, demonstrating a 5% relative improvement on three diverse real-world datasets (air quality, healthcare, anomaly detection).
• The method successfully aligns invariant sparse associative structures while leveraging variant domain-specific information.
• Visualization of learned structures provided insights into transferable knowledge, enhancing model transparency and interpretability.

Conclusions:

• The proposed SASA2 model effectively addresses the challenges of time-series unsupervised domain adaptation by focusing on stable associative structures.
• Integrating domain-specific information alongside invariant structures leads to significant performance gains.
• The method offers a more transparent and interpretable approach to domain adaptation in time-series analysis.