Advancing Imbalanced Domain Adaptation: Cluster-Level Discrepancy Minimization With a Comprehensive Benchmark.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) typically addresses covariate shift by aligning feature distributions.
• Standard UDA evaluation assumes invariant conditional label distributions, which is unrealistic for real-world data like long-tailed distributions.
• Real-world scenarios often involve both covariate shift and label shift, necessitating new adaptation strategies.

Purpose of the Study:

• Formulate imbalanced domain adaptation (IDA) for scenarios with both covariate and label shift.
• Develop a novel method to mitigate negative transfer caused by label shift in domain adaptation.
• Propose new evaluation protocols for realistic imbalanced domain adaptation scenarios.

Main Methods:

• Introduce cluster-level discrepancy minimization (CDM) for imbalanced domain adaptation.
• Employ cross-domain similarity learning to create discriminative clusters.
• Utilize learned clusters for both feature-level and distribution-level discrepancy minimization.

Main Results:

• Demonstrate that label shift can lead to negative transfer in standard domain adaptation.
• Show that CDM effectively alleviates the negative impact of label shift.
• Achieve significant improvements on imbalanced datasets like Office-31, Image-CLEF, and Office-Home.

Conclusions:

• Imbalanced domain adaptation (IDA) is crucial for realistic machine learning applications.
• CDM offers a robust solution for domain adaptation problems with imbalanced data.
• The proposed evaluation protocols and CDM method advance the field of domain adaptation.