Open Graph Benchmark: Datasets for Machine Learning on Graphs

proposal

Main paper can be found here and the code for it is found here.

• Provides diverse datasets from different domains
• Various sizes to test scalability
• Fully automated data loader, train-test split and evaluation methods for every dataset
• Different categories of tasks - node, link, graph property prediction
• Finally public leaderboard for inviting community contributions

ogb package

• provides utilities to download the datset and train/test split them.
• such utilities are provided for multiple datasets
• also provides utility methods to evaluate the resulting graph model too
• it is compatible with pytorch-geometric and DGL