OGB
Open Graph Benchmark — large-scale standardised graph ML benchmarks covering graph-level, node-level, and link-level prediction across diverse domains.
Sources: OGB.
ogbn-products, ogbg-molhiv, ogbn-papers100M, ogbg-molpcba and 12 more
Overview
Benchmark Groups
Benchmark groups are the sets of benchmarks that have been introduced together, e.g. the 5 LRGB datasets, all the OGB datasets, or the more recent GraphBench suite of benchmarks. Groups are ordered by the number of distinct papers reporting results for them.
MoleculeNet
MoleculeNet — benchmark collection for molecular machine learning.
Sources: MoleculeNet.
Heterogeneous Graph Benchmarks
Heterogeneous graph datasets from the MAGNN and HGB benchmark papers.
Sources: Heterogeneous Graph Benchmarks, Heterogeneous Graph Benchmarks.
LRGB
Long Range Graph Benchmark — five graph ML tasks designed to test long-range dependency modelling. Introduced by Dwivedi et al. (2022).
Sources: LRGB.
Peptides-struct, Peptides-func, PascalVOC-SP, PCQM-Contact and 1 more
Heterophily Benchmark
Five replacement datasets proposed for more reliable evaluation under graph heterophily.
Sources: Heterophily Benchmark.
Roman-empire, Amazon-ratings, Minesweeper, Tolokers and 1 more
GNNBenchmark
Benchmarking Graph Neural Networks — artificially and semi-artificially generated graph datasets.
Sources: GNNBenchmark.
Quantum Chemistry
Quantum chemistry molecular property prediction benchmarks.
Sources: Quantum Chemistry.
LINKX Benchmarks
Large-scale non-homophilous graph benchmarks from LINKX and related evaluations.
Sources: LINKX Benchmarks.
Penn94, arxiv-year, Pokec, snap-patents and 2 more
Knowledge Graphs
Knowledge graph completion and link prediction benchmarks.
Sources: Knowledge Graphs.
Other Graph Benchmarks
Miscellaneous graph learning datasets used across various papers: malware detection, 3D point clouds, text graphs, and large-scale graphs.
No single introduction paper.
Hypergraph Benchmarks
Hypergraph node classification benchmarks used in heterophily-aware and non-dissipative hypergraph GNN papers (AllSet, HyperGCN, etc.). Datasets span multiple source papers; AllSet is the partial group anchor.
Sources: Hypergraph Benchmarks.
GraphBench
GraphBench — a comprehensive graph learning benchmark covering diverse graph types and tasks spanning social networks, hardware design, logic, earth systems, and algorithmic reasoning.
Sources: GraphBench.
Algorithmic Reasoning, BA graph, ER graph, RB graph and 1 more
Algorithmic Graph Tasks
Algorithmic graph property prediction tasks (SSSP, eccentricity, etc.) used as benchmarks for expressive GNNs.
Sources: Algorithmic Graph Tasks.
DualEquiNet RNA Benchmarks
RNA 3D structure prediction benchmarks used in DualEquiNet (equivariant graph network for biological molecule geometry). Includes ribosome and synthetic RNA datasets.
Sources: DualEquiNet RNA Benchmarks.
ECHO Benchmark
ECHO — a benchmark for long-range graph propagation in physical simulation domains (synthetic, charge, and energy propagation tasks).
Sources: ECHO Benchmark.
City-Networks
City-Networks — a large-scale long-range benchmark on real-world city road graphs, introduced to test transductive node classification under very large graph diameters.
Sources: City-Networks.
LRIM Graph Benchmark
LRIM — the Long-Range Ising Model Graph Benchmark, a physics-grounded suite for controlled long-range graph learning evaluation across ten datasets spanning multiple graph sizes and difficulty levels.
Sources: Long-Range Ising Model Graph Benchmark.
Power Flow Benchmarks
Power grid test cases for scale-adaptive power flow analysis with graph neural networks. Cases correspond to standard IEEE power system test networks.
Sources: Power Flow Benchmarks.
HSG-12M
HSG-12M — a large-scale benchmark of Heterogeneous Spatial Graphs from real-world geospatial infrastructure data (transport, utilities, land-use).
Sources: HSG-12M.
Classic Legacy
Classic node/graph classification and regression benchmarks.
No single introduction paper.
Cora, CiteSeer, PubMed, Amazon-Photo and 8 more
TU Dortmund Legacy
TU Dortmund graph kernel benchmark datasets for graph classification.
Sources: TU Dortmund.
Heterophilic Graphs Legacy
Legacy heterophilic node classification benchmarks with known evaluation problems (WebKB, Wikipedia, and the five-class Actor graph).
Sources: Heterophilic Graphs, LINKX Benchmarks.