AGNN: Alternating Graph-Regularized Neural Networks to Alleviate Over-Smoothing

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Graph convolutional networks (GCNs) excel at analyzing graph-structured data but often suffer from over-smoothing, limiting network depth.
• Shallow GCNs restrict information propagation across complex graph datasets, hindering performance on intricate structures.
• Existing solutions like residual connections and model fusion have limitations in addressing deep layer embedding indistinguishability.

Purpose of the Study:

• To propose a novel alternating graph-regularized neural network (AGNN) to overcome the over-smoothing problem in GCNs.
• To enhance the propagation of information across remote neighbors in complex graph datasets.
• To generate more reliable predictions by improving feature distinguishability from different network layers.

Main Methods:

• Introduced an alternating graph-regularized neural network (AGNN) comprising graph convolutional layers (GCLs) and a graph embedding layer (GEL).
• Developed GEL from graph-regularized optimization with a Laplacian embedding term for periodic projection between feature spaces.
• Employed an improved Adaboost strategy to aggregate outputs from distinct layers, integrating multi-hop neighbor embeddings.

Main Results:

• The proposed GEL effectively alleviates the over-smoothing problem by projecting features between low-order and high-order spaces.
• AGNN generates more distinguishable features across different layers, leading to improved prediction reliability.
• Extensive experiments demonstrate AGNN's superior performance compared to state-of-the-art multilayer and multi-order graph neural networks.

Conclusions:

• AGNN offers a robust solution to the over-smoothing issue, enabling deeper and more effective GCNs.
• The model's ability to integrate multi-hop neighbor information through enhanced feature distinguishability is key to its success.
• AGNN represents a significant advancement in graph neural network architectures for complex graph-structural data analysis.