Critical Classes and Samples Discovering for Partial Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Partial domain adaptation (PDA) aims to transfer knowledge from labeled source to unlabeled target domains with fewer classes.
• Conventional PDA methods often struggle with inaccurate target predictions and uneven sample transferability, leading to negative transfer.
• Existing approaches overlook the importance of identifying critical source classes and ambiguous target samples.

Purpose of the Study:

• To propose a novel Critical Classes and Samples Discovering Network (CSDN) for more precise cross-domain alignment in partial domain adaptation.
• To address the limitations of inaccurate target predictions and varying sample transferability in current PDA methods.
• To enhance the performance and flexibility of transfer learning models in domain adaptation tasks.

Main Methods:

• CSDN employs an adaptive source class weighting scheme to dynamically select the most relevant source classes.
• A target ambiguous score is introduced to prioritize learning from uncertain target samples with high prediction inconsistency.
• The method utilizes co-training of two diverse classifiers for improved cross-domain distribution alignment.
• CSDN's weighting schemes are designed for easy integration with existing PDA and DA methods.

Main Results:

• CSDN effectively identifies critical source classes and target samples for improved adaptation.
• The proposed method demonstrates superior performance compared to state-of-the-art approaches on benchmark datasets.
• Experimental results validate the enhanced precision in cross-domain alignment achieved by CSDN.
• The flexibility of CSDN allows for performance boosts when coupled with other domain adaptation techniques.

Conclusions:

• CSDN offers a robust solution to the challenges of negative transfer and sample ambiguity in partial domain adaptation.
• The network significantly improves the accuracy and efficiency of knowledge transfer in domain adaptation.
• CSDN provides a flexible and effective framework that can be readily integrated into various transfer learning pipelines.