Correntropy-Induced Wasserstein GCN: Learning Graph Embedding via Domain Adaptation

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Graph embedding learns low-dimensional vertex representations from complex graph structures.
• Cross-domain graph embedding aims to transfer learned representations from a source to a target graph.
• Noise in graphs poses significant challenges for accurate information extraction and knowledge transfer.

Purpose of the Study:

• To develop a robust graph embedding architecture for noisy environments.
• To enhance the reliability of knowledge transfer in cross-graph embedding tasks.
• To improve the performance of graph analysis in the presence of graph noise.

Main Methods:

• A two-step correntropy-induced Wasserstein GCN (CW-GCN) architecture is proposed.
• The first step utilizes correntropy-induced loss to identify and exclude noisy nodes in the source graph.
• The second step employs Wasserstein distance to align the distributions of source and target graphs, facilitating knowledge transfer.

Main Results:

• CW-GCN demonstrates robustness in extracting helpful information from clean nodes.
• The method effectively mitigates the negative influence of noise during knowledge transfer.
• Extensive experiments show CW-GCN significantly outperforms state-of-the-art methods in noisy environments.

Conclusions:

• CW-GCN provides a robust solution for cross-graph embedding in the presence of noise.
• The proposed architecture facilitates reliable knowledge transfer for improved graph analysis.
• This approach offers a promising direction for handling noisy graph data.