Generalizing Aggregation Functions in GNNs: Building High Capacity and Robust GNNs via Nonlinear Aggregation

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning
• Artificial Intelligence

Background:

• GNNs utilize multi-layer architectures for nonlinear representation learning in graph data.
• Core GNN operation involves message propagation where nodes aggregate neighbor information.
• Existing GNNs commonly use linear (mean, sum) or max aggregators, facing limitations like over-smoothing, reduced capacity, and lack of detail sensitivity.

Purpose of the Study:

• To address limitations of existing aggregation methods in GNNs.
• To develop novel general nonlinear aggregators for improved message propagation.
• To enhance GNNs' nonlinearity, capacity, robustness, and detail sensitivity.

Main Methods:

• Re-thinking the message propagation mechanism in GNNs.
• Developing new general nonlinear aggregators for neighborhood information aggregation.
• Designing aggregators that optimally balance between max and mean/sum approaches.

Main Results:

• The proposed nonlinear aggregators provide an optimal balance between max and mean/sum.
• Achieved enhanced nonlinearity, improving network capacity and robustness.
• Demonstrated detail-sensitivity, preserving detailed node representation information.
• Experimental results confirm the effectiveness, high capacity, and robustness of the proposed methods.

Conclusions:

• The novel nonlinear aggregators overcome limitations of existing methods in GNNs.
• These aggregators enhance GNN performance by improving nonlinearity and detail awareness.
• The proposed approach offers a more effective and robust solution for graph learning tasks.