RAN-GNNs: Breaking the Capacity Limits of Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph neural networks (GNNs) are widely used for graph data analysis.
• Increasing GNN layers often yields diminishing performance returns.
• This limitation stems from the need to capture information from multiple neighborhood sizes simultaneously.

Purpose of the Study:

• To investigate randomly wired architectures as an alternative to deep stacking in GNNs.
• To demonstrate the effectiveness of randomly wired architectures in enhancing GNN capacity and representation quality.

Main Methods:

• Exploration of randomly wired architectures within the GNN framework.
• Theoretical analysis of randomly wired architectures behaving as ensembles of paths.
• Experimental validation across multiple tasks and graph convolution methods.

Main Results:

• Randomly wired architectures increase network capacity more effectively than deeper stacking.
• These architectures naturally merge contributions from varied receptive field sizes.
• Trainable weights allow modulation of receptive field width.

Conclusions:

• Randomly wired architectures provide a superior method for improving GNN performance.
• They offer a more effective strategy for learning richer node representations.
• This approach overcomes limitations associated with traditional deep GNNs.