Graph Decoupling Attention Markov Networks for Semisupervised Graph Node Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Graph neural networks (GNNs) are widely used for node classification.
• Standard GNNs aggregate neighbor information, but are susceptible to negative disturbances from edges connecting nodes with different labels.
• Existing attention mechanisms in GNNs often lack supervision and solely rely on feature similarity.

Purpose of the Study:

• To propose a decoupling attention mechanism for graph neural networks to address negative disturbances in node classification.
• To enhance the accuracy and robustness of GNNs by considering label dependency and refining graph structures.
• To improve information gain during message passing in GNNs.

Main Methods:

• Developed a decoupling attention mechanism that learns both hard and soft attention.
• Hard attention refines the graph structure by reducing inter-class edges, mitigating negative disturbances.
• Soft attention learns aggregation weights based on features over the refined graph, enhancing information gain.
• Formulated the model within the expectation-maximization (EM) framework, guiding label and feature propagation.

Main Results:

• The proposed method effectively reduces negative disturbances in graph node classification.
• Experimental results on six benchmark datasets demonstrate the superiority of the decoupling attention mechanism.
• The approach enhances information gain during message passing through refined graph structures and feature-based attention.

Conclusions:

• The decoupling attention mechanism offers a significant improvement over existing GNN methods for node classification.
• The EM-framework integration provides a principled way to learn attention for both label and feature propagation.
• This work advances GNNs by effectively handling label dependencies and improving robustness against noisy graph structures.