CALDA: Improving Multi-Source Time Series Domain Adaptation With Contrastive Adversarial Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Unsupervised Domain Adaptation (UDA) enhances machine learning in domains lacking labels by using labeled source domains.
• Weak supervision, using meta-domain information like label distributions, can further improve UDA performance.
• Multi-source UDA (MS-UDA) addresses complex scenarios with multiple labeled source domains.

Purpose of the Study:

• To propose a novel framework, CALDA, for robust multi-source unsupervised domain adaptation (MS-UDA) specifically for time series data.
• To synergistically combine contrastive and adversarial learning principles within CALDA.
• To leverage cross-source label information and weak supervision to boost performance.

Main Methods:

• CALDA employs adversarial learning to align feature representations between source and target domains.
• It utilizes contrastive learning to group similar labeled examples and separate dissimilar ones, reshaping the feature space.
• The framework uniquely integrates cross-source label information and weak supervision without requiring data augmentation or pseudo-labeling for time series.

Main Results:

• Empirical validation on human activity recognition, electromyography, and synthetic datasets demonstrated improved performance.
• Utilizing cross-source information significantly enhanced results compared to prior time series and contrastive adaptation methods.
• Weak supervision further boosted performance, even with noisy data, showcasing CALDA's robustness.

Conclusions:

• CALDA offers a generalizable strategy for MS-UDA in time series.
• The synergistic combination of contrastive and adversarial learning, along with cross-source label utilization, proves effective.
• The framework successfully addresses limitations of previous methods by avoiding data augmentation and pseudo-labeling.