Hierarchical Prototype Networks for Continual Graph Representation Learning.

Area of Science:

• Graph Representation Learning
• Continual Learning
• Machine Learning

Background:

• Continual learning in graph representation learning faces challenges with emerging node categories and catastrophic forgetting.
• Existing methods often overlook topological information or compromise plasticity for stability.

Purpose of the Study:

• To develop a novel method, Hierarchical Prototype Networks (HPNs), for continual graph representation learning that mitigates catastrophic forgetting.
• To represent continuously expanding graphs using hierarchical prototypes that capture abstract knowledge at different levels.

Main Methods:

• Leveraging Atomic Feature Extractors (AFEs) to encode node attributes and topological structure.
• Developing HPNs to adaptively select relevant AFEs and represent nodes with three levels of prototypes.
• Activating and refining only relevant AFEs and prototypes for new node categories to preserve performance on existing nodes.

Main Results:

• Demonstrated bounded memory consumption irrespective of the number of encountered tasks.
• Proved that learning new tasks does not alter prototypes matched to previous data under mild constraints, thus eliminating forgetting.
• Achieved superior performance over state-of-the-art baselines across five datasets with reduced memory consumption.

Conclusions:

• HPNs effectively address catastrophic forgetting in continual graph learning by employing a hierarchical prototype system.
• The proposed method offers a stable and plastic solution for evolving graph data, outperforming existing techniques.
• HPNs provide a memory-efficient and robust framework for dynamic graph representation learning.