Open Graph Benchmark: Datasets for Machine Learning on Graphs

The Open Graph Benchmark provides a diverse collection of challenging and realistic datasets designed to make graph machine learning research more scalable, robust, and reproducible. The datasets are large-scale, reaching over 100 million nodes and more than 1 billion edges, and span multiple important graph ML tasks across domains including social and information networks, biological networks, molecular graphs, source-code abstract syntax trees, and knowledge graphs. For each dataset, OGB supplies a unified evaluation protocol with meaningful application-specific data splits and evaluation metrics, along with an automated end-to-end pipeline that standardizes data loading, experimental setup, and model evaluation.

Extensive benchmark experiments across the datasets reveal that OGB poses significant challenges of scalability to large-scale graphs and of out-of-distribution generalization under realistic data splits, pointing to fruitful directions for future research. By publicly releasing datasets, data loaders, evaluation scripts, baseline code, and leaderboards, and committing to regular updates with community input, OGB established a standardized foundation for reproducible graph ML that addressed the field's prior reliance on small, inconsistent benchmarks.