A Graph Embedding Framework for Maximum Mean Discrepancy-Based Domain Adaptation Algorithms.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Domain adaptation addresses challenges where training and testing data distributions differ.
• Maximum Mean Discrepancy (MMD) is commonly used in domain adaptation to reduce distribution divergence.
• Existing MMD methods can reduce class compactness, leading to overlapping classes and poor classification.

Purpose of the Study:

• To propose a new graph embedding framework for Maximum Mean Discrepancy (MMD)-based domain adaptation.
• To enhance classification performance by ensuring discriminative features in cross-domain learning.
• To address the issue of reduced within-class compactness caused by traditional MMD minimization.

Main Methods:

• Developed a graph embedding framework incorporating intrinsic and penalty graphs.
• Revised the intrinsic graph to minimize within-class scatter, promoting feature discriminability.
• Proposed two strategies to instantiate the framework, resulting in four distinct models with penalty graphs.

Main Results:

• The proposed framework effectively minimizes within-class scatter, leading to more discriminative features.
• Experimental results on visual cross-domain datasets show significant improvements in classification performance.
• The new models outperform existing state-of-the-art domain adaptation methods.

Conclusions:

• The novel graph embedding framework offers a robust solution for MMD-based domain adaptation.
• Minimizing within-class scatter is crucial for achieving high classification accuracy in cross-domain scenarios.
• The proposed models provide enhanced discriminative capabilities for improved domain adaptation.