A cross-graph tuning-free GNN prompting framework

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Network (GNN) prompting aims to adapt models across tasks and graphs efficiently.
• Current GNN prompting methods often require task-specific parameter updates and struggle with cross-graph generalization.
• This limits performance and the practical utility of GNN prompting.

Purpose of the Study:

• To develop a novel Cross-graph Tuning-free Prompting Framework (CTP) for GNNs.
• To enable GNN models to adapt to new graphs without parameter tuning, facilitating plug-and-play inference.
• To support both homogeneous and heterogeneous graph structures.

Main Methods:

• Introduced the Cross-graph Tuning-free Prompting Framework (CTP).
• Designed CTP for seamless deployment on unseen graphs, eliminating the need for further parameter tuning.
• Validated CTP's effectiveness on few-shot prediction tasks across various graph types.

Main Results:

• CTP demonstrated significant performance improvements on few-shot prediction tasks.
• Achieved an average accuracy gain of 30.8% and a maximum gain of 54% compared to state-of-the-art methods.
• Showcased the framework's ability to generalize across different graphs and tasks.

Conclusions:

• CTP offers a robust and efficient solution for GNN adaptation across tasks and graphs.
• The tuning-free nature of CTP enables a plug-and-play GNN inference engine.
• This work provides a new perspective on prompt learning for GNNs, enhancing their practical applicability.