Contrastive Multi-View Representation Learning on Graphs

Kaveh Hassani, Amir Hosein Khas Ahmadi

Abstract

We introduce a self-supervised approach for learning node and graph level representations by contrasting structural views of graphs. We show that unlike visual representation learning, increasing the number of views to more than two or contrasting multi-scale encodings do not improve performance, and the best performance is achieved by contrasting encodings from first-order neighbors and a graph diffusion. We achieve new state-of-the-art results in self-supervised learning on 8 out of 8 node and graph classification benchmarks under the linear evaluation protocol. For example, on Cora (node) and Reddit-Binary (graph) classification benchmarks, we achieve 86.8% and 84.5% accuracy, which are 5.5% and 2.4% relative improvements over previous state-of-the-art. When compared to supervised baselines, our approach outperforms them in 4 out of 8 benchmarks. Source code is released at: this https URL

Benchmark Results

Tested on 4 datasets 1 baseline value reported too low

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

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
159	WL Trad baseline actual #159 vs paper #597 438 places worse	0.9040 ±0.0570 best reported	actual #159 vs paper #597 438 places worse paper 0.8070 ±0.0300 best at pub: Understanding Isomorphism Bias in Graph Data Sets (2019) best today: How Powerful are $K$-hop Message Passing Graph Neural Networks (NeurIPS 2022) paper 0.8070 ±0.0300 Δ at pub 0.0550 Δ today 0.0970
214	MI Hyb ★ proposed	0.8970 ±0.0110	This paper
257	InfoGraph GNN baseline actual #257 vs paper #272 15 places worse	0.8930 ±0.0514 best reported	actual #257 vs paper #272 15 places worse paper 0.8900 ±0.0110 best today: Improving Expressivity of GNNs with Subgraph-specific Factor Embedded Normalization (KDD 2023) Difference within reported uncertainty; not counted as underperformance. paper 0.8900 ±0.0110 Δ today 0.0030
274	GIN-ε GNN baseline actual #274 vs paper #274	0.8900 ±0.0600	actual #274 vs paper #274 paper 0.8900 ±0.0600 prev best: Fast Graph Representation Learning with PyTorch Geometric (ICLR 2019) Improvement within reported uncertainty; not highlighted. paper 0.8900 ±0.0600 Δ at pub 0.0560
332	MLG SVM baseline actual #332 vs paper #334 2 places worse	0.8794 ±0.0161 best reported	actual #332 vs paper #334 2 places worse paper 0.8790 ±0.0160 best at pub: The Multiscale Laplacian Graph Kernel (NeurIPS 2016) Difference within reported uncertainty; not counted as underperformance. paper 0.8790 ±0.0160 Δ at pub 0.0004
365	DGK SVM baseline actual #365 vs paper #366 1 place worse	0.8744 ±0.0272 best reported	actual #365 vs paper #366 1 place worse paper 0.8740 ±0.0270 best at pub: Provably Powerful Graph Networks (NeurIPS 2019) best today: Model-Driven Graph Contrastive Learning (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.8740 ±0.0270 Δ at pub 0.0004 Δ today 0.0004
403	MLP code ↗ Trad global	0.8660 ±0.0495	Breaking the Limits of Message Passing Graph Neural Networks code ↗
407	graph2vec Walk baseline actual #407 vs paper #515 108 places worse	0.8644 ±0.0125 best reported	actual #407 vs paper #515 108 places worse paper 0.8320 ±0.0960 best today: Graph Masked Autoencoders with Transformers (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.8320 ±0.0960 Δ today 0.0324
426	SP Trad baseline actual #426 vs paper #443 17 places worse	0.8580 ±0.0020 best reported	actual #426 vs paper #443 17 places worse paper 0.8520 ±0.0240 best at pub: Propagation Kernels (2014) Difference within reported uncertainty; not counted as underperformance. paper 0.8520 ±0.0240 Δ at pub 0.0060
472	Sub2Vec Walk baseline actual #472 vs paper #786 314 places worse	0.8447 ±0.0037 best reported	actual #472 vs paper #786 314 places worse paper 0.6110 ±0.1580 best today: Graph Masked Autoencoders with Transformers (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.6110 ±0.1580 Δ today 0.2337
500	random walk Walk baseline	0.8370 ±0.0150	This paper
506	GK SVM baseline actual #506 vs paper #575 69 places worse	0.8170 ±0.0210	actual #506 vs paper #575 69 places worse paper 0.8170 ±0.0210 prev best: Learning Convolutional Neural Networks for Graphs (ICML 2016) Improvement within reported uncertainty; not highlighted. paper 0.8170 ±0.0210 Δ at pub 0.0012
746	Node2Vec Walk baseline actual #746 vs paper #746	0.7263 ±0.1020 best reported	actual #746 vs paper #746 paper 0.7260 ±0.1020 best today: Model-Driven Graph Contrastive Learning (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.7260 ±0.1020 Δ today 0.0003

IMDB-BINARY

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	ID-GNN+BEC Hyb global	0.8150 ±0.0240	Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature
2	GIN+BEC Hyb global	0.8080 ±0.0330	Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature
3	GCN+BEC Hyb global	0.7990 ±0.0340	Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature
4	GCN GNN baseline actual #4 vs paper #56 52 places worse	0.7980 ±0.0230 best reported	actual #4 vs paper #56 52 places worse paper 0.7400 ±0.0340 best today: Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature (KDD 2025) paper 0.7400 ±0.0340 Δ today 0.0580
14	GIN GNN global	0.7810 ±0.0350	Depth-Adaptive Graph Neural Networks via Learnable Bakry-Émery Curvature
36	GIN-0 GNN baseline	0.7510 ±0.0510	This paper
48	GIN-ε GNN baseline	0.7430 ±0.0510	This paper
50	MI Hyb ★ proposed	0.7420 ±0.0070	This paper
56	GAT GNN baseline actual #56 vs paper #133 77 places worse	0.7407 ±0.0453 best reported	actual #56 vs paper #133 77 places worse paper 0.7050 ±0.0230 best today: Pooling Architecture Search for Graph Classification (2021) Difference within reported uncertainty; not counted as underperformance. paper 0.7050 ±0.0230 Δ today 0.0357
72	WL Trad baseline actual #72 vs paper #98 26 places worse	0.7340 ±0.0463 best reported	actual #72 vs paper #98 26 places worse paper 0.7230 ±0.0340 best today: Graph Learning with Distributional Edge Layouts (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7230 ±0.0340 Δ today 0.0110
82	InfoGraph GNN baseline actual #82 vs paper #82	0.7303 ±0.0087 best reported	actual #82 vs paper #82 paper 0.7300 ±0.0090 best today: Unifying Graph Contrastive Learning via Graph Message Augmentation (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7300 ±0.0090 Δ today 0.0003
99	GraphSAGE GNN baseline	0.7230 ±0.0530	This paper
125	graph2vec Walk baseline	0.7110 ±0.0050	This paper
140	DGK SVM baseline actual #140 vs paper #140	0.6700 ±0.0060	actual #140 vs paper #140 paper 0.6700 ±0.0060 prev best: Graph Capsule Convolutional Neural Networks (ICML 2018) Improvement within reported uncertainty; not highlighted. paper 0.6700 ±0.0060 Δ at pub 0.0004
141	MLG SVM baseline	0.6660 ±0.0030	This paper
142	GK SVM baseline actual #142 vs paper #142	0.6590 ±0.0100	actual #142 vs paper #142 paper 0.6590 ±0.0100 prev best: Graph Capsule Convolutional Neural Networks (ICML 2018) Improvement within reported uncertainty; not highlighted. paper 0.6590 ±0.0100 Δ at pub 0.0003
153	SP Trad baseline	0.5560 ±0.0020	This paper
154	Sub2Vec Walk baseline	0.5530 ±0.0150	This paper
157	random walk Walk baseline	0.5070 ±0.0030	This paper

IMDB-MULTI

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	GraphSNN GNN global	0.7790 ±0.0360	Empowering GNNs via Edge-Aware Weisfeiler-Leman Algorithm
2	SIN GNN global	0.7560 ±0.0320	Empowering GNNs via Edge-Aware Weisfeiler-Leman Algorithm
3	CIN GNN global	0.7560 ±0.0370	Empowering GNNs via Edge-Aware Weisfeiler-Leman Algorithm
5	GIN GNN global	0.7510 ±0.0510	Empowering GNNs via Edge-Aware Weisfeiler-Leman Algorithm
32	GIN-0 GNN baseline	0.5230 ±0.0280	This paper
36	GIN-ε GNN baseline	0.5210 ±0.0360	This paper
38	GCN GNN baseline	0.5190 ±0.0380	This paper
44	MI Hyb ★ proposed	0.5120 ±0.0050	This paper
46	GraphSAGE GNN baseline	0.5090 ±0.0220	This paper
53	graph2vec Walk baseline	0.5040 ±0.0090	This paper
67	InfoGraph GNN baseline	0.4970 ±0.0050	This paper
69	GAT GNN baseline actual #69 vs paper #90 21 places worse	0.4967 ±0.0430 best reported	actual #69 vs paper #90 21 places worse paper 0.4780 ±0.0310 best today: Pooling Architecture Search for Graph Classification (2021) Difference within reported uncertainty; not counted as underperformance. paper 0.4780 ±0.0310 Δ today 0.0187
72	WL Trad baseline actual #72 vs paper #93 21 places worse	0.4933 ±0.0475 best reported	actual #72 vs paper #93 21 places worse paper 0.4700 ±0.0050 best today: Graph Learning with Distributional Edge Layouts (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.4700 ±0.0050 Δ today 0.0233
109	DGK SVM baseline actual #109 vs paper #109	0.4460 ±0.0050	actual #109 vs paper #109 paper 0.4460 ±0.0050 prev best: Graph Capsule Convolutional Neural Networks (ICML 2018) Improvement within reported uncertainty; not highlighted. paper 0.4460 ±0.0050 Δ at pub 0.0005
110	GK SVM baseline actual #110 vs paper #110	0.4390 ±0.0040	actual #110 vs paper #110 paper 0.4390 ±0.0040 prev best: Graph Capsule Convolutional Neural Networks (ICML 2018) Improvement within reported uncertainty; not highlighted. paper 0.4390 ±0.0040 Δ at pub 0.0001
112	MLG SVM baseline	0.4120	This paper
113	SP Trad baseline	0.3800 ±0.0030	This paper
114	Sub2Vec Walk baseline	0.3670 ±0.0080	This paper
116	random walk Walk baseline	0.3470 ±0.0020	This paper

REDDIT-BINARY

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	GIN-0 GNN baseline	0.9240 ±0.0250	This paper
2	GIN-ε GNN baseline	0.9220 ±0.0230	This paper
3	DMoN GNN global	0.9100 ±0.0100	Torch Geometric Pool: the Pytorch library for pooling in Graph Neural Networks
33	GAT GNN baseline	0.8520 ±0.0330	This paper
36	MI Hyb ★ proposed	0.8450 ±0.0060	This paper
38	InfoGraph GNN baseline	0.8250 ±0.0140	This paper
42	DGK SVM baseline	0.7800 ±0.0040	This paper
43	GK SVM baseline	0.7730 ±0.0020	This paper
44	graph2vec Walk baseline	0.7580 ±0.0100	This paper
45	Sub2Vec Walk baseline	0.7150 ±0.0040	This paper
46	WL Trad baseline	0.6880 ±0.0040	This paper
47	SP Trad baseline	0.6410 ±0.0010	This paper
49	GCN GNN baseline actual #49 vs paper #50 1 place worse	0.5653 ±0.0260 best reported	actual #49 vs paper #50 1 place worse paper 0.5000 best today: Local Virtual Nodes for Alleviating Over-Squashing in Graph Neural Networks (2025) paper 0.5000 Δ today 0.0653

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