Fast Graph Representation Learning with PyTorch Geometric

Matthias Fey, J. E. Lenssen

Abstract

We introduce PyTorch Geometric, a library for deep learning on irregularly structured input data such as graphs, point clouds and manifolds, built upon PyTorch. In addition to general graph data structures and processing methods, it contains a variety of recently published methods from the domains of relational learning and 3D data processing. PyTorch Geometric achieves high data throughput by leveraging sparse GPU acceleration, by providing dedicated CUDA kernels and by introducing efficient mini-batch handling for input examples of different size. In this work, we present the library in detail and perform a comprehensive comparative study of the implemented methods in homogeneous evaluation scenarios.

Benchmark Results

Tested on 5 datasets 3 baseline values reported too low

Tables show presented results in global context (best 3 papers + GCN + MLP).

Classic

CiteSeer

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	CNA code ↗ GNN global	0.9575 ±0.0058	Graph Neural Networks Need Cluster-Normalize-Activate Modules code ↗
2	IS-GIB code ↗ Hyb global	0.9390 ±0.0187	Individual and Structural Graph Information Bottlenecks for Out-of-Distribution Generalization code ↗
3	EERM code ↗ GNN global	0.9112 ±0.0145	Individual and Structural Graph Information Bottlenecks for Out-of-Distribution Generalization code ↗
20	APPNP GNN baseline actual #20 vs paper #474 454 places worse	0.8142 ±0.0026 best reported	actual #20 vs paper #474 454 places worse paper 0.7180 ±0.0050 best today: Feature Expansion for Graph Neural Networks (ICML 2023) paper 0.7180 ±0.0050 Δ today 0.0962
35	GAT GNN baseline actual #35 vs paper #570 535 places worse	0.8080 ±0.0026 best reported	actual #35 vs paper #570 535 places worse paper 0.7080 ±0.0050 best at pub: Topology Adaptive Graph Convolutional Networks (2017) best today: Feature Expansion for Graph Neural Networks (ICML 2023) paper 0.7080 ±0.0050 Δ at pub 0.0170 Δ today 0.1000
44	GCN GNN baseline actual #44 vs paper #550 506 places worse	0.7110 ±0.0070	actual #44 vs paper #550 506 places worse paper 0.7110 ±0.0070 prev best: Topology Adaptive Graph Convolutional Networks (2017) best today: NTFormer: A Composite Node Tokenized Graph Transformer for Node Classification (2024) paper 0.7110 ±0.0070 Δ at pub 0.0080 Δ today 0.0923
61	GraphSAGE GNN global	0.7923 ±0.0053	Gophormer: Ego-Graph Transformer for Node Classification
96	SGC GNN baseline actual #96 vs paper #528 432 places worse	0.7765 ±0.0012 best reported	actual #96 vs paper #528 432 places worse paper 0.7130 ±0.0020 best at pub: Deep Learning on Graphs: A Survey (2018) best today: Node Representation Learning in Graph via Node-to-Neighbourhood Mutual Information Maximization (2022) paper 0.7130 ±0.0020 Δ at pub 0.0060 Δ today 0.0635
350	MLP code ↗ Trad global	0.7289 ±0.0010	Gumbel-MPNN: Graph Rewiring with Gumbel-Softmax code ↗
390	ARMA GNN baseline actual #390 vs paper #426 36 places worse	0.7250 best reported	actual #390 vs paper #426 36 places worse paper 0.7230 ±0.0110 best at pub: Graph Neural Networks With Convolutional ARMA Filters (2019) best today: GPNet: Simplifying Graph Neural Networks via Multi-channel Geometric Polynomials (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7230 ±0.0110 Δ at pub 0.0020 Δ today 0.0020
612	Cheby GNN baseline	0.7020 ±0.0100	This paper

PubMed

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	LGTL GT global	0.9518 ±0.0011	Adaptive Tokenization: On the Hop-Overpriority Problem in Tokenized Graph Learning Models
2	MixHop (LO) code ↗ GNN global	0.9513 ±0.0038	LOGIN: A Large Language Model Consulted Graph Neural Network Training Framework code ↗
3	GraphSAGE (LO) code ↗ GNN global	0.9511 ±0.0036	LOGIN: A Large Language Model Consulted Graph Neural Network Training Framework code ↗
17	APPNP GNN baseline actual #17 vs paper #439 422 places worse	0.8010 ±0.0020	actual #17 vs paper #439 422 places worse paper 0.8010 ±0.0020 prev best: Predict then Propagate: Graph Neural Networks meet Personalized PageRank (ICLR 2018) best today: LOGIN: A Large Language Model Consulted Graph Neural Network Training Framework (2024) Improvement within reported uncertainty; not highlighted. paper 0.8010 ±0.0020 Δ at pub 0.0037 Δ today 0.1345
18	GCN GNN baseline actual #18 vs paper #585 567 places worse	0.9296 ±0.0015 best reported	actual #18 vs paper #585 567 places worse paper 0.7900 ±0.0060 best at pub: Topology Adaptive Graph Convolutional Networks (2017) best today: Adaptive Tokenization: On the Hop-Overpriority Problem in Tokenized Graph Learning Models (2025) paper 0.7900 ±0.0060 Δ today 0.1396
20	GAT GNN baseline actual #20 vs paper #632 612 places worse	0.9233 ±0.0018 best reported	actual #20 vs paper #632 612 places worse paper 0.7850 ±0.0030 best at pub: Topology Adaptive Graph Convolutional Networks (2017) best today: Adaptive Tokenization: On the Hop-Overpriority Problem in Tokenized Graph Learning Models (2025) paper 0.7850 ±0.0030 Δ at pub 0.0050 Δ today 0.1383
116	SAGE GNN global	0.8881 ±0.0015	Efficient Text-Attributed Graph Learning through Selective Annotation and Graph Alignment
141	SGC GNN baseline actual #141 vs paper #602 461 places worse	0.8832 ±0.0055 best reported	actual #141 vs paper #602 461 places worse paper 0.7890 ±0.0010 best at pub: Deep Learning on Graphs: A Survey (2018) best today: LOGIN: A Large Language Model Consulted Graph Neural Network Training Framework (2024) paper 0.7890 ±0.0010 Δ today 0.0942
210	MLP code ↗ Trad global	0.8690 ±0.0074	On the Equivalence of Graph Convolution and Mixup code ↗
217	ARMA GNN baseline actual #217 vs paper #614 397 places worse	0.8678 ±0.0021 best reported	actual #217 vs paper #614 397 places worse paper 0.7880 ±0.0030 best at pub: Graph Neural Networks With Convolutional ARMA Filters (2019) best today: Graph Structure Prompt Learning: A Novel Methodology to Improve Performance of Graph Neural Networks (2024) paper 0.7880 ±0.0030 Δ at pub 0.0010 Δ today 0.0798
650	Cheby GNN baseline	0.7840 ±0.0040	This paper

TU Dortmund

MUTAG

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	MSH-GNN GNN global	0.9910 ±0.0030	Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing
2	ID-GNN+BEC Hyb global	0.9830 ±0.0360	Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature
3	SupCosine Hyb global	0.9830 ±0.0250	Supervised Contrastive Learning with Structure Inference for Graph Classification
39	GIN GNN global	0.9330 ±0.0290	ProtGNN: Towards Self-Explaining Graph Neural Networks
63	GCN GNN baseline actual #63 vs paper #719 656 places worse	0.9220 ±0.0440 best reported	actual #63 vs paper #719 656 places worse paper 0.7460 ±0.0770 best at pub: Graph Neural Networks With Convolutional ARMA Filters (2019) best today: Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature (KDD 2025) paper 0.7460 ±0.0770 Δ at pub 0.1110 Δ today 0.1760
190	SortPool GNN baseline actual #190 vs paper #674 484 places worse	0.9000 ±0.0224 best reported	actual #190 vs paper #674 484 places worse paper 0.7730 ±0.0890 best today: Graph Structure Prompt Learning: A Novel Methodology to Improve Performance of Graph Neural Networks (2024) paper 0.7730 ±0.0890 Δ today 0.1270
238	GIN-0 GNN baseline actual #238 vs paper #429 191 places worse	0.8940 ±0.0560 best reported	actual #238 vs paper #429 191 places worse paper 0.8570 ±0.0770 best today: InfoGCL: Information-Aware Graph Contrastive Learning (NeurIPS 2021) Difference within reported uncertainty; not counted as underperformance. paper 0.8570 ±0.0770 Δ today 0.0370
239	DiffPool GNN baseline actual #239 vs paper #447 208 places worse	0.8940 ±0.0460 best reported	actual #239 vs paper #447 208 places worse paper 0.8500 ±0.1030 best today: Graph Classification via Reference Distribution Learning: Theory and Practice (NeurIPS 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.8500 ±0.1030 Δ today 0.0440
274	GIN-ε GNN baseline actual #274 vs paper #511 237 places worse	0.8900 ±0.0600 best reported	actual #274 vs paper #511 237 places worse paper 0.8340 ±0.0750 best today: Prototypical Graph Contrastive Learning (2021) Difference within reported uncertainty; not counted as underperformance. paper 0.8340 ±0.0750 Δ today 0.0560
403	MLP code ↗ Trad global	0.8660 ±0.0495	Breaking the Limits of Message Passing Graph Neural Networks code ↗
593	Set2Set GNN baseline actual #593 vs paper #734 141 places worse	0.8084 ±0.0067 best reported	actual #593 vs paper #734 141 places worse paper 0.7370 ±0.0690 best today: SignGT: Signed Attention-based Graph Transformer for Graph Representation Learning (2023) paper 0.7370 ±0.0690 Δ today 0.0714
679	Graclus Trad baseline	0.7710 ±0.0720	This paper
689	top_k Trad baseline	0.7630 ±0.0750	This paper
718	SAGE GNN baseline	0.7490 ±0.0870	This paper
720	GlobalAttention GNN baseline	0.7460 ±0.0800	This paper

PROTEINS

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	MSH-GNN GNN global	0.9410 ±0.0330	Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing
2	GRDL GNN global	0.8260 ±0.0120	Graph Classification via Reference Distribution Learning: Theory and Practice
3	GMN code ↗ Hyb global	0.8225	Memory-Based Graph Networks code ↗
7	DiffPool GNN baseline actual #7 vs paper #235 228 places worse	0.8161 ±0.0232 best reported	actual #7 vs paper #235 228 places worse paper 0.7510 ±0.0350 best today: Graph Structure Prompt Learning: A Novel Methodology to Improve Performance of Graph Neural Networks (2024) paper 0.7510 ±0.0350 Δ today 0.0651
32	GIN GNN global	0.7880 ±0.0410	Permutation-Invariant graph partitioning: How graph neural networks capture structural …
67	SAGE GNN baseline actual #67 vs paper #324 257 places worse	0.7726 ±0.0228 best reported	actual #67 vs paper #324 257 places worse paper 0.7380 ±0.0360 best today: Semi-Supervised Hierarchical Graph Classification (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7380 ±0.0360 Δ today 0.0346
148	GCN GNN baseline actual #148 vs paper #360 212 places worse	0.7310 ±0.0380	actual #148 vs paper #360 212 places worse paper 0.7310 ±0.0380 prev best: Graph Neural Networks With Convolutional ARMA Filters (2019) Improvement within reported uncertainty; not highlighted. paper 0.7310 ±0.0380 Δ at pub 0.0210
149	MLP Trad global	0.7600 ±0.0320	Simplifying Graph Convolutional Networks with Redundancy-Free Neighbors
152	SortPool GNN baseline actual #152 vs paper #390 238 places worse	0.7598 ±0.0176 best reported	actual #152 vs paper #390 238 places worse paper 0.7240 ±0.0410 best today: Graph Structure Prompt Learning: A Novel Methodology to Improve Performance of Graph Neural Networks (2024) paper 0.7240 ±0.0410 Δ today 0.0358
301	Set2Set GNN baseline actual #301 vs paper #334 33 places worse	0.7429 best reported	actual #301 vs paper #334 33 places worse paper 0.7360 ±0.0370 best today: Memory-Based Graph Networks (ICLR 2020) Difference within reported uncertainty; not counted as underperformance. paper 0.7360 ±0.0370 Δ today 0.0069
367	Graclus Trad baseline	0.7300 ±0.0410	This paper
378	top_k Trad baseline	0.7270 ±0.0410	This paper
383	GIN-ε GNN baseline	0.7260 ±0.0490	This paper
387	GlobalAttention GNN baseline	0.7250 ±0.0450	This paper
406	GIN-0 GNN baseline	0.7210 ±0.0510	This paper

COLLAB

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	WL-MLP GT global	0.9792	1-WL Expressiveness Is (Almost) All You Need
2	MSH-GNN GNN global	0.9640 ±0.0070	Multi-Scale Harmonic Encoding for Feature-Wise Graph Message Passing
3	CoCN code ↗ GNN global	0.8722 ±0.0013	Scalable Graph Compressed Convolutions code ↗
23	GCN GNN baseline actual #23 vs paper #96 73 places worse	0.8260 ±0.0220 best reported	actual #23 vs paper #96 73 places worse paper 0.8060 ±0.0210 best today: Scalable Expressiveness through Preprocessed Graph Perturbations (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.8060 ±0.0210 Δ today 0.0200
30	GIN code ↗ GNN global	0.8220 ±0.0210	Scalable Expressiveness through Preprocessed Graph Perturbations code ↗
33	DiffPool GNN baseline actual #33 vs paper #161 128 places worse	0.8213 ±0.0043 best reported	actual #33 vs paper #161 128 places worse paper 0.7890 ±0.0230 best today: Scalable Graph Compressed Convolutions (2024) paper 0.7890 ±0.0230 Δ today 0.0323
107	Graclus Trad baseline actual #107 vs paper #140 33 places worse	0.8040 ±0.0150 best reported	actual #107 vs paper #140 33 places worse paper 0.7960 ±0.0200 best today: The expressive power of pooling in Graph Neural Networks (NeurIPS 2023) Difference within reported uncertainty; not counted as underperformance. paper 0.7960 ±0.0200 Δ today 0.0080
112	GIN-0 GNN baseline actual #112 vs paper #148 36 places worse	0.8020 ±0.0190 best reported	actual #112 vs paper #148 36 places worse paper 0.7930 ±0.0270 best today: InfoGCL: Information-Aware Graph Contrastive Learning (NeurIPS 2021) Difference within reported uncertainty; not counted as underperformance. paper 0.7930 ±0.0270 Δ today 0.0090
120	GIN-ε GNN baseline actual #120 vs paper #131 11 places worse	0.8010 ±0.0190 best reported	actual #120 vs paper #131 11 places worse paper 0.7980 ±0.0240 best today: Prototypical Graph Contrastive Learning (2021) Difference within reported uncertainty; not counted as underperformance. paper 0.7980 ±0.0240 Δ today 0.0030
125	SortPool GNN baseline actual #125 vs paper #187 62 places worse	0.8007 best reported	actual #125 vs paper #187 62 places worse paper 0.7770 ±0.0310 best today: CogDL: A Comprehensive Library for Graph Deep Learning (WWW 2021) paper 0.7770 ±0.0310 Δ today 0.0237
137	SAGE GNN baseline	0.7970 ±0.0170	This paper
138	top_k Trad baseline	0.7970 ±0.0220	This paper
144	GlobalAttention GNN baseline	0.7960 ±0.0220	This paper
142	Set2Set GNN baseline	0.7960 ±0.0230	This paper
151	MLP Trad global	0.7920 ±0.0240	Simplifying Graph Convolutional Networks with Redundancy-Free Neighbors

SOTA performance (not significantly different from the best) Baseline / comparison (this paper) Global context (other papers) Best at pub — reported result worse than the best known when published (best reported in green) Improved at pub — later paper surpassed this result Best today — OK at publication; today's leaderboard best is elsewhere