Graph Memory Learning: Imitating Lifelong Remembering and Forgetting of Brain Networks.

Area of Science:

• Graph Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Real-world graph data changes rapidly, posing challenges for existing graph models.
• Frequent retraining of graph models is computationally expensive and impractical.
• Existing models struggle with continuous data influx and data withdrawal.

Purpose of the Study:

• To introduce a novel concept of graph memory learning for dynamic graph data.
• To develop a brain-inspired graph memory learning framework (BGML) for efficient knowledge management.
• To enable graph models to selectively remember new knowledge while forgetting outdated information.

Main Methods:

• Proposed Brain-inspired Graph Memory Learning (BGML) framework.
• Incorporated a multi-granular hierarchical progressive learning mechanism for feature graph grain learning.
• Introduced an information self-assessment ownership mechanism for incremental data integrity.

Main Results:

• BGML effectively mitigates the conflict between memorization and forgetting in graph memory learning.
• The framework enables multi-level perception of local details in evolving graphs.
• Extensive experiments on node classification datasets confirmed BGML's excellent performance across various tasks.

Conclusions:

• BGML offers an efficient solution for handling dynamic graph data.
• The proposed mechanisms enhance the model's ability to adapt to new information while preserving past knowledge.
• BGML demonstrates superior performance in managing evolving graph structures and information.