On Target Shift in Adversarial Domain Adaptation.

Area of Science:

• Machine Learning
• Computer Vision
• Domain Adaptation

Background:

• Domain discrepancy hinders machine learning generalization.
• Label shift, differing class distributions between domains, is understudied.
• Behavioral studies often exhibit natural label shift.

Purpose of the Study:

• To propose a novel method, Domain Adversarial nets for Target Shift (DATS), for domain-invariant representation learning under label shift.
• To address the challenge of differing class proportions between training and testing domains.
• To extend the framework for multi-domain adaptation.

Main Methods:

• Utilizing adversarial deep learning for domain-invariant feature representation.
• Employing distribution matching to estimate label proportions in unlabeled target datasets.
• Developing a source domain upweighting scheme for multi-domain scenarios.

Main Results:

• DATS effectively learns domain-invariant representations even with significant label shift.
• The method demonstrates strong performance in both synthetic and real-world experiments.
• Empirical results validate the practical importance of addressing label shift.

Conclusions:

• DATS provides a robust solution for the prevalent issue of label shift in machine learning.
• The proposed approach enhances model generalization across domains with varying label distributions.
• This work highlights the significance of targeted methods for specific domain adaptation challenges.