Co-Embedding of Nodes and Edges With Graph Neural Networks

Area of Science:

• Machine Learning
• Graph Theory
• Data Science

Background:

• Graphs are crucial data structures in diverse fields like social networks and biology.
• Graph embedding transforms complex graph data into a usable format for machine learning.
• Existing Graph Convolutional Networks (GCNs) primarily focus on node features, neglecting valuable edge information.

Purpose of the Study:

• To develop a general graph embedding framework that incorporates both node and edge features.
• To address the limitations of current GCNs in utilizing edge information.
• To improve the performance of machine learning tasks on graph-structured data.

Main Methods:

• Introduced CensNet (Convolution with Edge-Node Switching), a novel graph neural network.
• Utilized the line graph of the original undirected graph to switch node and edge roles.
• Proposed two new graph convolution operations for effective feature propagation.

Main Results:

• CensNet successfully embeds both nodes and edges into a latent feature space.
• Achieved state-of-the-art or competitive performance across four graph learning tasks.
• Demonstrated effectiveness on academic citation networks and quantum chemistry graphs.

Conclusions:

• CensNet offers a powerful approach for graph learning tasks involving both node and edge features.
• The edge-node switching mechanism enhances information extraction from graph data.
• This framework advances the capabilities of graph neural networks in real-world applications.