Learning Target-Domain-Specific Classifier for Partial Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised Domain Adaptation (UDA) typically assumes identical label spaces between source and target domains.
• This assumption is often violated in real-world applications, necessitating more flexible adaptation methods.
• Partial Domain Adaptation (PDA) addresses scenarios where the target domain's label space is a subset of the source domain's.

Purpose of the Study:

• To address the challenge of negative transfer in Partial Domain Adaptation (PDA).
• To propose a novel method, Target-Domain-Specific Classifier learning-based Domain Adaptation (TSCDA), for improved knowledge transfer.
• To enhance the discriminative power of classifiers in the target domain.

Main Methods:

• TSCDA employs a soft-weighed Maximum Mean Discrepancy (MMD) criterion for partial feature distribution alignment.
• It incorporates target-domain-specific classifier learning using pseudo-labels and auxiliary classifiers.
• A peers-assisted learning module minimizes prediction discrepancies among target-specific classifiers.

Main Results:

• TSCDA significantly alleviates negative transfer caused by source outliers.
• The method demonstrates superior performance compared to state-of-the-art approaches on benchmark PDA datasets.
• Average performance gains of 4% on Office-31 and 5.6% on Office-Home were observed.

Conclusions:

• TSCDA offers a robust solution for Partial Domain Adaptation by effectively handling differing label spaces.
• The proposed method enhances classifier discriminability and reduces negative transfer.
• TSCDA represents a significant advancement in unsupervised domain adaptation research.