Hybrid graph attention learning with pseudo-label guided adaptive evolution

Area of Science:

• Graph Representation Learning
• Machine Learning
• Artificial Intelligence

Background:

• Existing Graph Attention Network (GAT) methods focus on node embedding-level attention, neglecting crucial graph topology information.
• This limitation leads to challenges in distinguishing node importance, semantic deviation in predictions, and insufficient interaction between labeled and unlabeled nodes.

Purpose of the Study:

• To propose a hybrid graph attention learning mechanism integrating both node embedding-level and structure embedding-level attentions.
• To introduce a dynamic graph evolution mechanism for adaptive graph structure correction and improved robustness.
• To enhance semantic alignment between graph representations and label predictions.

Main Methods:

• Developed a hybrid attention mechanism combining node embedding-level and structure embedding-level attentions.
• Implemented a dynamic graph evolution strategy involving topology pruning and node mixing guided by pseudo-labels.
• Created a closed-loop framework for representation learning and graph optimization through adaptive feature and structure mixing.

Main Results:

• The proposed method achieves more comprehensive and accurate modeling of node neighboring relationships.
• Demonstrated enhanced robustness to noisy graphs and improved semantic alignment for label prediction.
• Achieved significant performance improvements over existing baselines on real-world graph datasets.

Conclusions:

• The hybrid graph attention learning mechanism offers superior performance in exploring accurate attention and discriminative representation learning.
• The dynamic graph evolution mechanism effectively corrects graph structures and enhances model robustness.
• The integrated approach leads to significant advancements in graph representation learning and prediction accuracy.