Graph Alignment for Benchmarking Graph Neural Networks and Learning Positional Encodings

This article introduces a new benchmarking methodology for graph neural networks (GNNs) centered on the graph alignment problem, which seeks to maximize overlapping edges between unlabeled graphs. The authors present techniques for generating graph alignment datasets of varying difficulty, demonstrating that anisotropic models outperform isotropic ones in structure-only tasks. Additionally, the study reveals that node embeddings from self-supervised GNN pre-training can serve as effective positional encodings for transformers, achieving 98% accuracy in graph structure reconstruction, and provides an open-source Python package for dataset generation and benchmarking.