DeeperGCN: Training Deeper GCNs With Generalized Aggregation Functions.

Area of Science:

• Machine Learning
• Graph Representation Learning

Background:

• Graph Neural Networks (GNNs) are increasingly used for representation learning on graph-structured data.
• Recent advancements enable training very deep GNNs, showing success in areas like point cloud learning and protein interaction prediction.
• The performance of deep GNNs on large-scale graphs remains an area for investigation.

Purpose of the Study:

• To investigate the impact of aggregation function choice on the performance of deep GNNs in large-scale graph settings.
• To address the sensitivity of GNNs to standard aggregation functions (mean, max, sum) across diverse datasets.
• To propose a novel, learnable class of aggregation functions for enhanced deep GNN performance.

Main Methods:

• Systematic study of aggregation function effects on deep GNNs.
• Introduction of Generalized Aggregation Functions, a new class of permutation-invariant and differentiable functions.
• End-to-end learning of Generalized Aggregation Function parameters tailored to specific tasks.

Main Results:

• GNNs exhibit significant sensitivity to the choice of aggregation functions.
• Generalized Aggregation Functions extend existing methods, offering greater flexibility and adaptability.
• Deep residual GNNs equipped with Generalized Aggregation Functions achieve state-of-the-art results on Open Graph Benchmark (OGB) datasets.

Conclusions:

• The selection of aggregation functions is critical for deep GNN performance, especially in large-scale scenarios.
• Generalized Aggregation Functions provide a robust and adaptable solution to improve deep GNNs.
• This work advances deep learning on graphs by offering a more effective approach to aggregation.