Weighted Correlation Embedding Learning for Domain Adaptation.

Area of Science:

• Computer Science
• Machine Learning

Background:

• Domain adaptation is crucial for applying knowledge from source to target domains, particularly in pattern recognition and image classification.
• Existing methods often focus on minimizing distribution differences, potentially leading to negative transfer by ignoring task-specific knowledge and distribution outliers.

Purpose of the Study:

• To propose a novel domain adaptation method, Weighted Correlation Embedding Learning (WCEL), for improved image classification.
• To address the limitations of existing methods by considering task-specific knowledge transfer and mitigating negative transfer.

Main Methods:

• WCEL integrates correlation learning, graph embedding, and sample reweighting into a unified model.
• It extracts maximum correlated features and utilizes two graphs to preserve discriminant and neighborhood information.
• An efficient sample reweighting strategy is employed to handle varying confidence levels in target domain prediction.

Main Results:

• Extensive experiments on benchmark databases demonstrate the effectiveness of WCEL.
• The proposed method shows superior performance compared to existing state-of-the-art domain adaptation algorithms in image classification tasks.

Conclusions:

• WCEL offers a robust approach to domain adaptation for image classification.
• The method successfully addresses negative transfer and improves knowledge transfer efficiency.