Datasets
Browse current and legacy datasets by benchmark group; linked dataset pages include result tables and SOTA progression plots.
OGB
Sources: OGBogbn-products
Predict one of 47 top-level product categories for every product node using its description representation and co-purchase neighborhood.
Best: Jacobi Accuracy: 0.9
ogbg-molhiv
Predict whether a compound inhibits HIV replication.
Best: Multi-RF Fusion with Multi-GNN Blending ROC-AUC: 0.85
ogbn-papers100M
Predict a primary arXiv subject area for the approximately 1.5 million labeled arXiv papers while using the full 111-million-node citation network.
Best: SGC Accuracy: 0.79
ogbl-collab
Rank future author-author collaboration edges above sampled non-edges, given collaborations observed in earlier years.
Best: PROXI Hits@50: 0.77
ogbn-arxiv
Predict each paper's primary arXiv computer-science category from its content representation and citation context.
Best: GIANT-XRT Accuracy: 0.76
ogbl-ddi
Rank true drug-drug interaction edges above sampled non-interacting pairs.
Best: ELGNN Hits@20: 0.98
ogbl-ppa
Predict held-out biologically meaningful protein-protein association edges.
Best: OGB Leader Hits@100: 0.65
ogbl-citation2
For each 2019 source paper, rank a withheld true reference above 1,000 older, uncited candidate papers.
Best: CFG MRR: 0.9
ogbn-proteins
Predict all applicable Gene Ontology-derived function labels for each protein.
Best: LD ROC-AUC: 0.89
ogbn-mag
Predict the conference or journal of each paper using paper content, citations, authors, affiliations, and fields of study.
Best: LMSPS+LP+MS Accuracy: 0.58
PCQM4Mv2
Predict each molecule's HOMO-LUMO energy gap, a continuous quantum-chemical property calculated with density functional theory.
Best: DeMol MAE: 0.06
ogbg-ppa
Given a protein-association neighborhood, predict the broad taxonomic group of its source species.
Best: A-NLSF Accuracy: 0.81
ogbg-code2
Predict the subtokens in a Python method's name from the method body represented as an abstract syntax tree.
Best: GMN F1: 0.21
ogbl-biokg
Predict missing biomedical triples by ranking the true head or tail above sampled replacements of the same entity type.
Best: PairRE+LRE MRR: 0.84
MoleculeNet
Sources: MoleculeNetTox21
Multi-task toxicity prediction (12 tasks) on NCI Tox21 challenge compounds.
Best: HIMP ROC-AUC: 0.87
BACE
Binary classification — BACE-1 inhibitory activity (binding results for human beta-secretase 1).
Best: ProtoMol ROC-AUC: 0.91
BBBP
Binary classification — blood-brain barrier permeability of molecules.
Best: MAGPrompt+ ROC-AUC: 0.94
ClinTox
Clinical toxicity prediction — FDA-approved vs. failed drugs (2 tasks).
Best: SE(3)-I MPPH ROC-AUC: 0.99
ToxCast
Multi-task toxicological profile prediction (617 tasks) from the EPA ToxCast program.
Best: GeomGCL ROC-AUC: 0.76
Lipophilicity
Octanol/water distribution coefficient (logD) regression on 4,200 molecules.
Best: MemGNN RMSE: 0.56
Heterogeneous Graph Benchmarks
Sources: Heterogeneous Graph Benchmarks, Heterogeneous Graph BenchmarksLRGB
Sources: LRGBPeptides-struct
Predict 11 graph-level geometric descriptors computed from each peptide's 3D structure.
Best: GCN MAE: 0.24
Peptides-func
Predict 10 non-exclusive peptide-function labels from OGB-featured molecular graphs without hydrogens.
Best: S^2GCN AP: 0.73
PascalVOC-SP
Assign one of 21 semantic-segmentation labels, including background, to every superpixel node.
Best: NeuralWalker F1: 0.49
PCQM-Contact
Rank atom pairs that are more than five covalent-bond hops apart but less than 3.5 angstrom apart in the molecule's computed 3D geometry.
Best: FloydNet MRR: 0.61
Heterophily Benchmark
Sources: Heterophily BenchmarkRoman-empire
Wikipedia Roman Empire article node classification (18 syntactic roles) from the heterophilic benchmark.
Best: Dir-Poly Accuracy: 0.95
Amazon-ratings
Amazon product co-purchase node classification (5 rating classes) from the heterophilic benchmark.
Best: UniGAP Accuracy: 0.64
Minesweeper
Synthetic Minesweeper grid graph binary node classification from the heterophilic benchmark.
Best: BuNN ROC-AUC: 0.99
Tolokers
Yandex Toloka worker network binary node classification from the heterophilic benchmark.
Best: Polynormer-r ROC-AUC: 0.86
GNNBenchmark
Sources: GNNBenchmarkMNIST
Graph classification on MNIST superpixel graphs (10 digit classes, 70k graphs).
Best: Schrödinger Accuracy: 0.99
CIFAR10
Graph classification on CIFAR10 superpixel graphs (10 classes, 60k graphs).
Best: RandAlign Accuracy: 0.79
CLUSTER
Semi-supervised node classification on stochastic block model graphs (6 classes).
Best: GRPE-Small Accuracy: 0.82
PATTERN
Semi-supervised node classification — detect specific graph patterns (2 classes).
Best: DeepGraph (48) Accuracy: 0.91
Quantum Chemistry
Sources: Quantum ChemistryLINKX Benchmarks
Sources: LINKX BenchmarksPenn94
Facebook100 Penn94 social network node classification (binary gender).
Best: DC-GNN Accuracy: 0.87
arxiv-year
OGB-arXiv node classification by publication year (5 classes, heterophilic variant).
Best: LargeScaleNet Accuracy: 0.66
Pokec
Slovak social network Pokec node classification (binary gender, 1.6M nodes).
Best: SAGE Accuracy: 0.86
snap-patents
USPTO patent citation graph node classification (5 time-based classes).
Best: FaberNet Accuracy: 0.75
twitch-gamers
Twitch streamer network binary node classification (streaming language).
Best: DGT Accuracy: 0.66
Knowledge Graphs
Sources: Knowledge GraphsWN18RR
Knowledge graph link prediction on WordNet (40,943 entities, 11 relations).
Best: FedSage MRR: 0.84
FB15k-237
Knowledge graph link prediction on Freebase (14,541 entities, 237 relations).
Best: sum, Msg_r^3, f(t)=0 MRR: 0.85
Other Graph Benchmarks
No single introduction paperAmazon2M
Predict each product’s category on a roughly two-million-node co-purchase graph.
Best: MA-GCL Accuracy: 0.89
MalNet-Tiny
Predict one of five balanced malware families; nodes have no attributes on Android function-call graphs.
Best: GMN Accuracy: 0.94
ShapeNet-Part
Predict a part label for every point across 16 object categories and 50 part classes on 3D point clouds.
Best: GPS + k-MIP F1: 0.83
MovieLens Film
Predict one or more of nine genre labels per film; this is not the Actor/Film dataset on a MovieLens/IMDb movie graph.
Best: CLN-HWN-mini Micro-F1: 0.57
Hypergraph Benchmarks
Sources: Hypergraph BenchmarksUAT
US Airport Transactions hypergraph node classification (4 region classes, 1190 nodes).
Best: SCAGC F1: 0.73
BAT
Brazilian Airport Transactions hypergraph node classification (4 region classes, 131 nodes).
Best: GraphLearner F1: 0.75
EAT
European Airport Transactions hypergraph node classification (4 region classes, 399 nodes).
Best: SCAGC F1: 0.73
House
US House of Representatives voting hypergraph node classification (2 classes).
Best: HealHGNN Accuracy: 0.77
Cora-CA
Cora co-authorship hypergraph node classification (7 subject classes).
Best: HAN (full batch) Accuracy: 0.84
DBLP-CA
DBLP co-authorship hypergraph node classification (research area).
Best: HealHGNN Accuracy: 0.92
Walmart
Walmart co-purchase trip hypergraph node classification (product category, 11 classes).
Best: HealHGNN Accuracy: 0.68
GraphBench
Sources: GraphBenchAlgorithmic Reasoning
Predict task-specific node, edge, or graph outputs for seven classical graph algorithms.
Best: Naive multi-task Accuracy: 1
BA graph
Predict optimal objectives or decode feasible MIS, max-cut, and coloring solutions on Barabási–Albert graphs.
Best: Diffusion Wavelets Accuracy: 0.94
ER graph
Predict optimal objectives or decode feasible MIS, max-cut, and coloring solutions on Erdős–Rényi graphs.
Best: Diffusion Wavelets Accuracy: 0.93
RB graph
Predict optimal objectives or decode feasible MIS, max-cut, and coloring solutions on RB-model graphs.
Best: Diffusion Wavelets Accuracy: 0.85
Algorithmic Graph Tasks
Sources: Algorithmic Graph TasksSSSP
Predict every node’s shortest-path distance from a marked source on synthetic graphs.
Best: TANGO-GPS MSE: -5.5
Eccentricity
Predict each node’s maximum shortest-path distance to any other node on synthetic graphs.
Best: GCN-SSM + k-hop MSE: -4.3
DualEquiNet RNA Benchmarks
Sources: DualEquiNet RNA BenchmarksECHO Benchmark
Sources: ECHO BenchmarkECHO-Synth
Predict graph diameter, node eccentricity, and single-source shortest-path distances on synthetic graphs.
Best: GHR MAE: 0.035
City-Networks
Sources: City-NetworksParis
Predict an accessibility class for every intersection on the Paris road graph.
Best: GAT Accuracy: 0.56
Shanghai
Predict an accessibility class for every intersection on the Shanghai road graph.
Best: GAT Accuracy: 0.73
London
Predict an accessibility class for every intersection on the London road graph.
Best: GatedGCN Accuracy: 0.61
Los Angeles
Predict an accessibility class for every intersection on the Los Angeles road graph.
Best: Exphormer Accuracy: 0.64
LRIM Graph Benchmark
Sources: Long-Range Ising Model Graph BenchmarkPower Flow Benchmarks
Sources: Power Flow BenchmarksHSG-12M
Sources: HSG-12MHSG-12M
Classify each Hamiltonian spectral graph into one of 1,401 characteristic-polynomial classes.
Best: GraphSAGE Accuracy: 0.55
HSG-topology
Classify each topology-deduplicated Hamiltonian spectral graph into one of 1,401 characteristic-polynomial classes.
Best: GraphSAGE Accuracy: 0.62
Classic
Legacy No single introduction paperCora
Predict one of seven research-topic labels for every paper in a single citation network.
Best: GraphSAGE+CNA Accuracy: 0.94
CiteSeer
Predict each paper’s research topic from bag-of-words features and citation links on one citation graph.
Best: CNA Accuracy: 0.96
PubMed
Predict one of three diabetes-related paper classes from TF-IDF features and citations on one citation graph.
Best: LGTL Accuracy: 0.95
Amazon-Photo
Predict each product’s category from review-derived features and product links on one co-purchase graph.
Best: GraphTARIF Accuracy: 0.97
Amazon-Computers
Predict each product’s category from review-derived features and product links on one co-purchase graph.
Best: GraphTARIF Accuracy: 0.95
Predict the subreddit/community of each post from text-derived features on a post-interaction graph.
Best: EnGCN F1: 0.97
Coauthor-Physics
Predict each author’s physics field from paper-keyword features on one coauthorship graph.
Best: GrokFormer Accuracy: 0.98
PPI
Predict 121 non-exclusive Gene Ontology labels per protein from sequence/role features across protein-interaction graphs.
Best: GAMLP(JK) F1: 1
WikiCS
Predict each article’s computer-science category from text-derived features on a Wikipedia hyperlink graph.
Best: GCN Accuracy: 0.93
Coauthor-CS
Predict each author’s computer-science field from paper-keyword features on one coauthorship graph.
Best: DoG Accuracy: 0.97
MUTAG
Graph classification of mutagenic aromatic and heteroaromatic nitro compounds (2 classes).
Best: MSH-GNN Accuracy: 0.99
NCI1
Graph classification of chemical compounds screened for anti-cancer activity.
Best: WL-MLP Accuracy: 1
PROTEINS
Graph classification of proteins as enzyme or non-enzyme (2 classes).
Best: MSH-GNN Accuracy: 0.94
COLLAB
Graph classification of scientific collaboration networks (3 classes).
Best: WL-MLP Accuracy: 0.98
ENZYMES
TU Dortmund graph classification on 6 enzyme reaction classes (600 graphs, subset of BRENDA database).
Best: GMN Accuracy: 0.79
D&D
Graph classification of proteins into enzymes and non-enzymes via D&D representation.
Best: GC-GNN (GCN) Accuracy: 0.92
IMDB-BINARY
Binary graph classification of movie collaboration networks (action vs. romance).
Best: ID-GNN+BEC Accuracy: 0.81
RDT-B
Classify Reddit threads as question-answering or discussion; this is an alias of REDDIT-BINARY.
Best: OEPG Accuracy: 0.96
IMDB-MULTI
Multi-class graph classification of movie collaboration networks (3 genres).
Best: GraphSNN Accuracy: 0.78
REDDIT-BINARY
Binary graph classification of Reddit discussion threads (question-answering vs. online discussion).
Best: GIN-0 Accuracy: 0.92
REDDIT-MULTI-5K
Multi-class graph classification of Reddit discussion threads (5 subreddits).
Best: SpeGCL Accuracy: 0.59
Texas
Classify each webpage as student, project, course, staff, or faculty using the page text and hyperlink graph.
Best: P^2GNN Accuracy: 0.97
Cornell
WebKB node classification on Cornell university webpages (5 classes).
Best: P^2GNN Accuracy: 0.95
Actor
Actor co-occurrence node classification (5 classes) from the Geom-GCN paper.
Best: DiGGR Accuracy: 0.45
Chameleon
Wikipedia Chameleon page node classification (5 classes, heterophilic).
Best: Hetero-S (GAT 8-layer) Accuracy: 0.87