Ricci Curvature-Based Graph Sparsification for Continual Graph Representation Learning.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Artificial Intelligence

Background:

• Continual learning aims to learn new tasks without forgetting previous ones.
• Memory replay is effective for Euclidean data but struggles with graph data's topological nature.
• Existing graph learning methods often neglect edge information during memory replay.

Purpose of the Study:

• To develop a novel memory replay technique for continual graph representation learning.
• To address the limitations of existing methods in capturing topological information in graphs.
• To improve the performance and efficiency of graph neural networks in continual learning scenarios.

Main Methods:

• Proposed Subgraph Episodic Memory (SEM) to store topological information as computation subgraphs.
• Utilized Ricci curvature for graph sparsification, identifying informative nodes and edges.
• Developed a computationally efficient surrogate for Ricci curvature for large-scale graph applicability.

Main Results:

• SEM significantly outperforms state-of-the-art approaches on four public datasets.
• Achieved strong performance in the challenging class incremental learning (class-IL) setting.
• Demonstrated comparable performance to joint training in class-IL, a significant advancement.

Conclusions:

• SEM effectively preserves and replays critical topological information for continual graph learning.
• The Ricci curvature-based sparsification enhances memory efficiency and learning performance.
• SEM offers a robust solution for both task-IL and class-IL settings, advancing graph-based continual learning.