TL;DR
This paper introduces CALDA, a novel framework for Multi-Source Unsupervised Domain Adaptation (MS-UDA) in time series data. CALDA combines adversarial and contrastive learning to effectively leverage multiple labeled source domains for unlabeled target domains, and it can further benefit from weak supervision signals from the target domain. This approach significantly improves performance over existing methods, especially in complex real-world scenarios.
What Problem Does This Paper Address?
Machine learning models often struggle when applied to new datasets (target domains) that differ from the data they were trained on (source domains), especially when labels are unavailable for the target domain. This is known as the domain adaptation problem. When multiple labeled source domains are available, Multi-Source Unsupervised Domain Adaptation (MS-UDA) aims to improve performance on the target domain. However, existing MS-UDA methods for time series data often fail to fully exploit the rich information available across source domains or to effectively incorporate weak supervision signals (like label distributions) from the target domain. This paper addresses the challenge of building robust MS-UDA models for time series data by synergistically combining adversarial learning for domain alignment and contrastive learning for leveraging cross-source label information, while also incorporating weak supervision.
Methods
CALDA is built upon three core components:
- Multi-Source Domain Adversarial Training: Similar to prior work, CALDA uses a domain adversary to align feature representations from different source domains and the target domain. This encourages the feature extractor to produce domain-invariant features, making the model more transferable.
- Contrastive Learning: This is a key novelty. CALDA employs a supervised contrastive loss to pull together examples with the same label across different source domains and push apart examples with different labels. This helps to learn more discriminative representations by explicitly leveraging cross-source label information.
- Weak Supervision: CALDA can optionally incorporate target domain label proportions (e.g., the expected frequency of each activity) to further guide the learning process. This is achieved through a regularization term that aligns the model's predicted class distribution with the provided weak supervision.
Crucially, CALDA avoids common pitfalls in time series adaptation, such as requiring extensive data augmentation or relying on potentially unreliable pseudo-labeling of the target domain.
Datasets
The authors evaluated CALDA on a variety of datasets:
- Synthetic Datasets:
- SW (Sine Wave): Generated by summing two sine waves, with inter-domain or intra-domain translation/rotation shifts to create domain differences.
- GMM (Gaussian Mixture Model): Multivariate (9-dimensional) data generated from mixtures of Gaussians, with varying complexity (1GMM, 2GMM, 3GMM) to simulate different levels of domain shift.
- Real-World Datasets:
- Human Activity Recognition (HAR): UCI HAR, UCI HHAR, WISDM AR, WISDM AT.
- Electromyography (EMG): Myo EMG, NinaPro Myo.
Key Insights and Ideas
- Cross-Source Information is Crucial: Leveraging label information across different source domains (not just within each domain) significantly boosts performance in MS-UDA for time series. This is a key differentiator from methods that only consider within-source relationships.
- Contrastive Learning for Time Series MS-UDA: Applying contrastive learning principles to MS-UDA for time series is effective, especially when it focuses on aligning representations from different source domains.
- Weak Supervision Adds Value: Even with noisy label proportions, weak supervision provides a valuable signal that further improves adaptation performance, particularly on unbalanced datasets.
- Avoiding Pseudo-Labeling: For time series, relying on pseudo-labeling the target domain can be detrimental due to large domain shifts. CALDA's ability to perform contrastive learning solely on source domains is a significant advantage.
- Adversarial Learning Remains Essential: The adversarial component is vital for achieving domain-invariant features and good transferability, even when contrastive learning is employed.
Key Results
- CALDA significantly outperforms baseline methods (No Adaptation, CAN, CoDATS) on both synthetic and real-world datasets, achieving improvements of up to 20.1% on WISDM AR compared to CAN.
- CALDA-Any,R and CALDA-XS,H were identified as the most promising instantiations, showing strong performance across various datasets.
- The inclusion of weak supervision further boosted performance, with CALDA-WS showing improvements of up to 16.5% over CoDATS-WS on synthetic datasets and demonstrating robustness to noise in the weak supervision signals.
- Experiments showed that CALDA can handle varying time series lengths due to its use of global average pooling.
- The adversarial component was shown to be critical, with methods excluding it performing significantly worse.
Limitations and Future Work
- Pseudo-Labeling Sensitivity: While CALDA avoids pseudo-labeling for contrastive learning, the paper notes that determining which pseudo-labeling techniques work well for time series remains an open question for future work.
- Hard vs. Random Sampling: The benefit of hard sampling over random sampling for contrastive learning was found to be dataset-dependent, suggesting further research is needed to understand this interaction.
- Unevenly Spaced Time Series: The current framework assumes uniform time intervals. Future work could explore adapting CALDA for unevenly spaced time series data, such as electronic health records.
- Data Augmentation for Time Series: Investigating data augmentation techniques compatible with time series data for domain adaptation is suggested.
- Cross-Modality Adaptation: Extending CALDA to cross-modality adaptation is another promising future direction.
Summary updated