Graph Batch Coarsening framework for scalable graph neural networks.

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning on Graphs
• Scalable Graph Analytics

Background:

• Scaling graph neural networks (GNNs) to large graphs is challenging due to the neighborhood explosion phenomenon.
• Existing sampling-based mini-batch methods (node-wise, layer-wise, subgraph sampling) incur overhead and yield inconsistent performance.
• Random sampling in GNN training can be inefficient and impact model effectiveness.

Purpose of the Study:

• To introduce Graph Batch Coarsening (GBC), a novel framework for scalable GNN training.
• To provide a general solution that facilitates the training of arbitrary GNN models on large graphs.
• To overcome the limitations of random sampling in GNN training.

Main Methods:

• Graph Batch Coarsening (GBC) preprocesses input graphs into smaller subgraphs for mini-batch training.
• The framework employs a graph decomposition method leveraging label propagation.
• A novel graph coarsening algorithm specifically designed for GNN training is utilized.

Main Results:

• GBC completely avoids random sampling, simplifying the training process.
• The framework requires no modifications to existing GNN models or their hyperparameters.
• Empirical results show superior performance in accuracy, reduced training time, and lower memory usage across various graph scales.

Conclusions:

• Graph Batch Coarsening (GBC) presents an effective and generalizable approach for scalable GNN training.
• The method significantly improves efficiency and performance compared to traditional sampling techniques.
• GBC offers a promising direction for applying GNNs to large-scale graph data.