Towards Fair Knowledge Transfer for Imbalanced Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain adaptation (DA) addresses data scarcity by leveraging source domain knowledge.
• Existing DA methods often overlook fairness issues with imbalanced source data, hindering minority class adaptation.
• Cross-domain learning faces challenges due to domain shift and data imbalance.

Purpose of the Study:

• To propose a Towards Fair Knowledge Transfer (TFKT) framework for imbalanced cross-domain learning.
• To enhance the adaptation of under-represented minority classes in domain adaptation.
• To mitigate domain shift while ensuring fairness in knowledge transfer.

Main Methods:

• Developed a novel cross-domain knowledge propagation technique using structure graphs.
• Implemented a cross-domain fulfillment augmentation strategy for domain adaptation.
• Utilized hybrid distinct classifiers and cross-domain prototype alignment for robust classification and domain shift mitigation.

Main Results:

• The TFKT framework effectively addresses fairness challenges in imbalanced cross-domain learning.
• Significant improvements in adapting minority source sets were observed.
• The proposed model achieved over 20% improvement in overall accuracy on two benchmarks compared to state-of-the-art DA models.

Conclusions:

• The TFKT framework provides a unified approach to tackle both fairness and domain shift in cross-domain learning.
• The proposed methods demonstrate superior performance, particularly for minority classes in imbalanced scenarios.
• This work advances fair and effective domain adaptation techniques.