Ricci Curvature-Based Semi-Supervised Learning on an Attributed Network.

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning
• Network Science

Background:

• Traditional neural networks struggle with graph-structured data.
• Graph Convolutional Networks (GCNs) extend convolution to graphs but use Laplacian matrices for feature aggregation.
• Graph data is inherently non-Euclidean, necessitating non-Euclidean mathematical tools for analysis.

Purpose of the Study:

• To introduce a novel semi-supervised learning model, the Ricci curvature-based graph convolutional neural network (RCGCN).
• To leverage Riemannian geometry and Ricci curvature for improved feature aggregation in graph neural networks.
• To enhance the understanding of geometric structures in network data.

Main Methods:

• The study treats networks as discrete manifolds.
• Ricci curvature is employed to assign importance weights to neighboring nodes during feature aggregation.
• The aggregation pattern is inspired by GCN but utilizes Ricci curvature for more nuanced weighting.

Main Results:

• The proposed RCGCN model demonstrates a linear scaling with the number of edges.
• Ricci curvature effectively captures the geometric structure and relationships within the network.
• RCGCN achieved significant performance gains over baseline methods on benchmark datasets.

Conclusions:

• RCGCN offers a novel and effective approach to semi-supervised learning on graph data.
• The use of Ricci curvature provides a superior method for node importance weighting compared to traditional Laplacian methods.
• The model shows promise for analyzing complex network structures by incorporating geometric insights.