Co-embedding of edges and nodes with deep graph convolutional neural networks

Area of Science:

• Graph Neural Networks
• Deep Learning
• Machine Learning

Background:

• Graph neural networks (GNNs) excel with non-Euclidean data but often have shallow structures limiting information flow.
• Existing GNNs typically treat node and edge feature learning as separate tasks, hindering comprehensive feature extraction.

Purpose of the Study:

• To propose a novel message-passing framework for constructing deep GNNs comparable to deep convolutional neural networks (CNNs).
• To develop a framework for simultaneously learning node and edge embeddings, improving feature representation.
• To introduce the Co-embedding of Edges and Nodes with Deep Graph Convolutional Neural Networks (CEN-DGCNN) model.

Main Methods:

• Developed a novel message-passing framework integrating node and multi-dimensional edge features.
• Proposed a deep graph convolutional neural network model that prevents over-smoothing and captures long-distance dependencies.
• Introduced a graph convolutional layer for simultaneous learning of node and multi-dimensional edge embeddings using an attention mechanism.
• Implemented a multi-dimensional edge feature encoding method and constructed multi-channel filters for node information processing.

Main Results:

• The proposed CEN-DGCNN model effectively integrates node and edge features within a deep architecture.
• The model successfully extracts non-local structural and refined high-order node features by capturing long-distance dependencies.
• Extensive experiments demonstrated that CEN-DGCNN significantly outperforms existing GNN baseline methods.

Conclusions:

• CEN-DGCNN offers an effective approach to building deep GNNs with enhanced information transmission capabilities.
• Simultaneous co-embedding of node and edge features leads to superior performance in graph representation learning.
• The proposed framework addresses key limitations of existing GNN models, paving the way for more powerful graph analysis tools.