Decouple Graph Neural Networks: Train Multiple Simple GNNs Simultaneously Instead of One.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) face efficiency challenges due to exponential node dependency growth with increased layers.
• This limits stochastic optimization, making GNN training time-consuming.

Purpose of the Study:

• To develop a more efficient training framework for GNNs.
• To address the limitations of traditional GNN training methods.

Main Methods:

• Decoupling multi-layer GNNs into simple modules for efficient training.
• Implementing classical forward training (FT) and a novel backward training (BT) mechanism.
• Utilizing stochastic optimization algorithms for module training.

Main Results:

• The proposed framework enables efficient training of GNN modules without information distortion.
• Backward training allows modules to perceive information from deeper layers, enhancing training.
• Theoretical analysis shows minimal error accumulation in linear modules for unsupervised tasks.

Conclusions:

• The novel framework significantly improves GNN training efficiency while maintaining reasonable performance.
• This decoupled approach offers a promising direction for future GNN research and application.