SEA++: Multi-Graph-Based Higher-Order Sensor Alignment for Multivariate Time-Series Unsupervised Domain Adaptation.

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• Unsupervised Domain Adaptation (UDA) methods reduce label dependency by minimizing domain discrepancies.
• Existing UDA methods struggle with Multivariate Time-Series (MTS) data due to sensor-level distribution variations.
• This limitation hinders the effectiveness of UDA for MTS data, which is common in real-world applications.

Purpose of the Study:

• To address the challenges of UDA for MTS data by proposing a new framework called Multivariate Time-Series Unsupervised Domain Adaptation (MTS-UDA).
• To introduce Sensor Alignment (SEA) to tackle domain discrepancy at both local and global sensor levels.
• To enhance the proposed method to SEA++ by incorporating advanced alignment techniques.

Main Methods:

• Proposed SEnsor Alignment (SEA) framework for MTS-UDA.
• Developed endo-feature alignment to align sensor features and their correlations at the local sensor level.
• Designed exo-feature alignment to enforce restrictions on global sensor features for global sensor level alignment.
• Extended SEA to SEA++ with multi-graph-based higher-order alignment for enhanced feature and correlation alignment.

Main Results:

• SEA and SEA++ demonstrate state-of-the-art performance on six public MTS datasets.
• The proposed methods effectively address domain discrepancy at both local and global sensor levels.
• Empirical results validate the superiority of SEA and SEA++ for MTS-UDA tasks.

Conclusions:

• SEA and SEA++ offer effective solutions for Multivariate Time-Series Unsupervised Domain Adaptation.
• The sensor-level alignment approach significantly improves UDA performance on MTS data.
• The proposed methods advance the field of domain adaptation for complex time-series data.