Joint Clustering and Discriminative Feature Alignment for Unsupervised Domain Adaptation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised Domain Adaptation (UDA) seeks to classify target data using labeled source data from related but different distributions.
• Existing UDA methods often match marginal distributions, neglecting discriminative target features and cross-domain feature alignment, leading to suboptimal results.

Purpose of the Study:

• To address limitations in UDA by proposing a novel approach that simultaneously mines discriminative features and aligns class-discriminative features.
• To improve the performance of UDA models by integrating feature mining and alignment within a unified framework.

Main Methods:

• The Joint Clustering and Discriminative Feature Alignment (JCDFA) approach unifies discriminative feature mining and class-discriminative feature alignment.
• JCDFA jointly learns shared representations for supervised source data classification and discriminative target data clustering.
• Cross-domain alignment is achieved through semi-supervised contrastive learning and conditional Maximum Mean Discrepancy (MMD) to minimize intra-class and maximize inter-class compactness.

Main Results:

• Experiments on benchmarks like Office-31, ImageCLEF-DA, Office-Home, and VisDA-C show JCDFA significantly outperforms state-of-the-art UDA methods.
• Ablation studies confirm the effectiveness of individual components and the synergistic benefit of combining feature mining and alignment strategies.

Conclusions:

• The proposed JCDFA framework effectively integrates discriminative feature mining and cross-domain alignment for improved Unsupervised Domain Adaptation.
• The cooperative learning perspective enhances model generalization by leveraging both source and target domain information more effectively.