A Novel Composite Graph Neural Network

Area of Science:

• Machine Learning
• Graph Learning
• Artificial Intelligence

Background:

• Graph neural networks (GNNs) excel at processing graph-structured data but struggle with real-world noisy or undefined graph structures.
• Graph learning offers solutions for handling these challenges, prompting the development of more robust GNN methods.

Purpose of the Study:

• To develop a novel approach, composite GNN (C-GNN), to enhance the robustness of GNNs for semisupervised node classification.
• To characterize both sample and feature relations within a unified graph structure for improved performance.

Main Methods:

• Introduced composite graphs (C-graphs) that unify sample similarities and tree-based feature importance graphs.
• Developed a joint learning framework for multiaspect C-graphs and neural network parameters.
• Evaluated performance through experiments on nine benchmark datasets, comparing C-GNNs against variants focusing solely on sample or feature relations.

Main Results:

• The proposed C-GNN method achieved superior performance in semisupervised node classification across most benchmark datasets.
• The method demonstrated significant robustness against feature noises.
• Experimental results validated the effectiveness of unifying both sample and feature relations.

Conclusions:

• Composite GNNs offer a robust and high-performing solution for semisupervised node classification, particularly in the presence of noisy or incomplete graph structures.
• The unified C-graph approach effectively models both inter-sample similarities and intra-sample feature importance.
• This work advances graph learning by providing a method to improve GNN performance and resilience in complex, real-world scenarios.