BiKT: Unleashing the Potential of GNNs via Bi-Directional Knowledge Transfer

Area of Science:

• Graph Neural Networks (GNNs)
• Representation Learning
• Machine Learning

Background:

• Message-passing paradigm is key to GNNs, with focus on feature propagation.
• Feature transformation in GNNs is underexplored.
• GNNs may not fully utilize inherent feature transformation capabilities.

Purpose of the Study:

• Investigate GNN feature transformation performance.
• Propose Bi-directional Knowledge Transfer (BiKT) to enhance GNNs.
• Unleash potential of feature transformation operations.

Main Methods:

• Empirical investigation of feature transformation in GNNs.
• Development of BiKT as a plug-and-play approach.
• Derived representation learning model shares parameters with original GNN.
• Bi-directional knowledge injection between GNN and derived model.

Main Results:

• BiKT improves GNN performance by 0.5%-4% across 7 datasets and 5 GNNs.
• The derived model shows competitive or superior performance to the original GNN.
• Theoretical analysis confirms BiKT enhances generalization bounds via domain adaptation.

Conclusions:

• BiKT effectively boosts GNN performance by optimizing feature transformation.
• The derived model offers a powerful, independently applicable tool for downstream tasks.
• BiKT provides a flexible and architecture-agnostic enhancement for GNNs.