A deep graph convolutional neural network architecture for graph classification.

Area of Science:

• Graph Neural Networks
• Deep Learning
• Computer Science

Background:

• Graph Convolutional Networks (GCNs) excel with non-Euclidean data but are typically shallow (3-4 layers).
• Shallow GCNs struggle with high-level feature extraction due to over-smoothing and localized filters.
• Existing GCN architectures are limited in depth, hindering performance on complex graph tasks.

Purpose of the Study:

• To propose a novel framework enabling deeper GCNs and mitigating over-smoothing.
• To introduce a new spatial convolution layer for multiscale feature extraction.
• To develop and evaluate a deep GCN model for graph classification.

Main Methods:

• Introduced the Non-local Message Passing (NLMP) framework for flexible, deep GCN design.
• Developed a novel spatial graph convolution layer for multiscale node feature extraction.
• Designed and implemented Deep Graph Convolutional Neural Network II (DGCNNII), a 32-layer model for graph classification.

Main Results:

• DGCNNII effectively suppresses the over-smoothing phenomenon in deep GCNs.
• The proposed NLMP framework allows for the flexible design of very deep graph convolutional networks.
• Experiments demonstrated that DGCNNII outperforms numerous shallow GCN baselines on benchmark datasets.

Conclusions:

• The proposed NLMP framework and spatial convolution layer enable the creation of effective deep GCNs.
• DGCNNII achieves superior performance in graph classification tasks compared to shallow GCN models.
• This work advances the capability of GCNs for complex graph representation learning.