Memory-Based Graph Networks

Amir Hosein Khas Ahmadi, Kaveh Hassani, Parsa Moradi, Leo J. Lee, Q. Morris

Abstract

Graph neural networks (GNNs) are a class of deep models that operate on data with arbitrary topology represented as graphs. We introduce an efficient memory layer for GNNs that can jointly learn node representations and coarsen the graph. We also introduce two new networks based on this layer: memory-based GNN (MemGNN) and graph memory network (GMN) that can learn hierarchical graph representations. The experimental results show that the proposed models achieve state-of-the-art results in eight out of nine graph classification and regression benchmarks. We also show that the learned representations could correspond to chemical features in the molecule data. Code and reference implementations are released at: https://github.com/amirkhas/GraphMemoryNet

Benchmark Results

Tested on 7 datasets 6 baseline values reported too low

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

TU Dortmund

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 Hyb ★ proposed	0.8225	This paper
7	DiffPool GNN baseline actual #7 vs paper #132 125 places worse	0.7625	actual #7 vs paper #132 125 places worse paper 0.7625 prev best: Fast Graph Representation Learning with PyTorch Geometric (ICLR 2019) best today: Graph Structure Prompt Learning: A Novel Methodology to Improve Performance of Graph Neural Networks (2024) Improvement within reported uncertainty; not highlighted. paper 0.7625 Δ at pub 0.0115 Δ today 0.0536
8	MemGNN Hyb baseline	0.8135	This paper
32	GIN GNN global	0.7880 ±0.0410	Permutation-Invariant graph partitioning: How graph neural networks capture structural …
58	TopKPool GNN baseline	0.7768	This paper
113	ShortestPath SVM baseline	0.7643	This paper
149	MLP Trad global	0.7600 ±0.0320	Simplifying Graph Convolutional Networks with Redundancy-Free Neighbors
152	SortPool GNN baseline actual #152 vs paper #190 38 places worse	0.7554	actual #152 vs paper #190 38 places worse paper 0.7554 prev best: ASAP: Adaptive Structure Aware Pooling for Learning Hierarchical Graph Representations (AAAI 2019) paper 0.7554 Δ at pub 0.0163
154	SAGPool GNN baseline actual #154 vs paper #416 262 places worse	0.7595 ±0.0452 best reported	actual #154 vs paper #416 262 places worse paper 0.7186 best today: Embedding Graphs on Grassmann Manifold (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.7186 Δ today 0.0409
156	PatchySan GNN baseline actual #156 vs paper #245 89 places worse	0.7590 ±0.0280 best reported	actual #156 vs paper #245 89 places worse paper 0.7500 best today: Simplifying Graph Convolutional Networks with Redundancy-Free Neighbors (2025) Difference within reported uncertainty; not counted as underperformance. paper 0.7500 Δ today 0.0090
155	GraphSAGE GNN baseline actual #155 vs paper #453 298 places worse	0.7590 ±0.0320 best reported	actual #155 vs paper #453 298 places worse paper 0.7048 best at pub: Graph Convolutional Networks with EigenPooling (KDD 2019) best today: Simplifying Graph Convolutional Networks with Redundancy-Free Neighbors (2025) paper 0.7048 Δ at pub 0.0412 Δ today 0.0542
190	WL Trad baseline actual #190 vs paper #325 135 places worse	0.7560 ±0.0040 best reported	actual #190 vs paper #325 135 places worse paper 0.7376 best at pub: On Valid Optimal Assignment Kernels and Applications to Graph Classification (NeurIPS 2016) paper 0.7376 Δ at pub 0.0184
196	Sparse HGC GNN baseline	0.7546	This paper
222	WL Optimal SVM baseline	0.7526	This paper
301	Set2Set GNN baseline actual #301 vs paper #301	0.7429	actual #301 vs paper #301 paper 0.7429 prev best: Fast Graph Representation Learning with PyTorch Geometric (ICLR 2019) Improvement within reported uncertainty; not highlighted. paper 0.7429 Δ at pub 0.0069
370	Graphlet Trad baseline	0.7291	This paper
380	ECC GNN baseline	0.7265	This paper
384	CliquePool GNN baseline	0.7259	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 code ↗ GNN global	0.8260 ±0.0220	Scalable Expressiveness through Preprocessed Graph Perturbations code ↗
57	TopKPool GNN baseline actual #57 vs paper #191 134 places worse	0.7756	actual #57 vs paper #191 134 places worse paper 0.7756 prev best: Edge Contraction Pooling for Graph Neural Networks (2019) best today: Embedding Graphs on Grassmann Manifold (2022) paper 0.7756 Δ at pub 0.1296 Δ today 0.0384
92	WL Optimal SVM baseline	0.8074	This paper
117	GMN Hyb ★ proposed	0.8018	This paper
125	SortPool GNN baseline actual #125 vs paper #269 144 places worse	0.8007 best reported	actual #125 vs paper #269 144 places worse paper 0.7376 best at pub: Fast Graph Representation Learning with PyTorch Geometric (ICLR 2019) best today: CogDL: A Comprehensive Library for Graph Deep Learning (WWW 2021) paper 0.7376 Δ at pub 0.0394 Δ today 0.0631
136	GraphSAGE GNN baseline actual #136 vs paper #308 172 places worse	0.7970 ±0.0170 best reported	actual #136 vs paper #308 172 places worse paper 0.6825 best today: SignGT: Signed Attention-based Graph Transformer for Graph Representation Learning (2023) paper 0.6825 Δ today 0.1145
142	Set2Set GNN baseline actual #142 vs paper #286 144 places worse	0.7960 ±0.0230 best reported	actual #142 vs paper #286 144 places worse paper 0.7175 best at pub: Fast Graph Representation Learning with PyTorch Geometric (ICLR 2019) paper 0.7175 Δ at pub 0.0785
151	MLP Trad global	0.7920 ±0.0240	Simplifying Graph Convolutional Networks with Redundancy-Free Neighbors
153	WL Trad baseline actual #153 vs paper #166 13 places worse	0.7910 ±0.0010 best reported	actual #153 vs paper #166 13 places worse paper 0.7861 best at pub: On Valid Optimal Assignment Kernels and Applications to Graph Classification (NeurIPS 2016) paper 0.7861 Δ at pub 0.0049
209	MemGNN Hyb baseline	0.7700	This paper
242	Sparse HGC GNN baseline	0.7546	This paper
257	CliquePool GNN baseline	0.7450	This paper
280	PatchySan GNN baseline	0.7260	This paper
312	ECC GNN baseline	0.6779	This paper
322	Graphlet Trad baseline	0.6466	This paper
331	ShortestPath SVM baseline	0.5910	This paper

ENZYMES

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	GMN Hyb ★ proposed	0.7866	This paper
2	MemGNN Hyb baseline	0.7550	This paper
3	TFGW SP Hyb global	0.7510 ±0.0500	Template based Graph Neural Network with Optimal Transport Distances
17	GatedGCN code ↗ GNN global	0.7000 ±0.0494	GNNBenchmark code ↗
18	WL Trad baseline actual #18 vs paper #143 125 places worse	0.6950 ±0.0320 best reported	actual #18 vs paper #143 125 places worse paper 0.5343 best at pub: Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on Graphs (2017) best today: Template based Graph Neural Network with Optimal Transport Distances (NeurIPS 2022) paper 0.5343 Δ at pub 0.0562 Δ today 0.1607
26	GraphSAGE GNN baseline actual #26 vs paper #137 111 places worse	0.6817 ±0.0545 best reported	actual #26 vs paper #137 111 places worse paper 0.5425 best at pub: Graph Neural Networks With Convolutional ARMA Filters (2019) best today: GNNBenchmark (JMLR 2023) paper 0.5425 Δ at pub 0.0605 Δ today 0.1392
32	GIN code ↗ GNN global	0.6767 ±0.0583	GNNBenchmark code ↗
58	Sparse HGC GNN baseline	0.6417	This paper
70	DiffPool GNN baseline actual #70 vs paper #90 20 places worse	0.6053	actual #70 vs paper #90 20 places worse paper 0.6053 prev best: Hierarchical Representation Learning in Graph Neural Networks with Node Decimation Pooling (2019) best today: MathNet: Haar-Like Wavelet Multiresolution-Analysis for Graph Representation and Learning (2020) paper 0.6053 Δ at pub 0.3593 Δ today 0.0200
88	CliquePool GNN baseline	0.6071	This paper
93	Set2Set GNN baseline actual #93 vs paper #93	0.6015	actual #93 vs paper #93 paper 0.6015 prev best: Graph Convolutional Networks with EigenPooling (KDD 2019) paper 0.6015 Δ at pub 0.2215
94	WL Optimal SVM baseline	0.6013	This paper
120	SortPool GNN baseline	0.5712	This paper
124	MLP code ↗ Trad global	0.5583 ±0.0352	GNNBenchmark code ↗
128	Graphlet Trad baseline actual #128 vs paper #184 56 places worse	0.4103	actual #128 vs paper #184 56 places worse paper 0.4103 prev best: The Multiscale Laplacian Graph Kernel (NeurIPS 2016) best today: Contrastive Graph Few-Shot Learning (2022) paper 0.4103 Δ at pub 0.3008 Δ today 0.1427
182	ShortestPath SVM baseline	0.4232	This paper

MoleculeNet

Tox21

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	HIMP Hyb global	0.8740 ±0.0050	Graph convolutions that can finally model local structure
2	GINE w/ VN GNN global	0.8720 ±0.0030	Graph convolutions that can finally model local structure
3	NaiveGINe+ K=3 w/ VN GNN global	0.8700 ±0.0040	Graph convolutions that can finally model local structure
14	GCN GNN baseline actual #14 vs paper #23 9 places worse	0.8290 ±0.0060	actual #14 vs paper #23 9 places worse paper 0.8290 ±0.0060 prev best: Adaptive Graph Convolutional Neural Networks (AAAI 2018) best today: Graph convolutions that can finally model local structure (2020) paper 0.8290 ±0.0060 Δ at pub 0.0809 Δ today 0.0110
24	GIN GNN global	0.8290 ±0.0069	Towards Interpretable Sparse Graph Representation Learning with Laplacian Pooling
26	MemGNN Hyb ★ proposed	0.8280 ±0.0040	This paper
32	Kernel SVM SVM baseline	0.8220 ±0.0060	This paper
33	Weave GNN baseline	0.8200 ±0.0100	This paper
41	Bypass Trad baseline	0.8100 ±0.0130	This paper
50	Multitask GNN baseline	0.8030 ±0.0120	This paper
63	IRV Trad baseline	0.7990 ±0.0060	This paper
76	Logistic Regression Trad baseline	0.7940 ±0.0150	This paper
77	XGBoost Boost baseline	0.7940 ±0.0140	This paper
156	Random Forest RF baseline	0.7690 ±0.0150	This paper

BACE

Global leaderboard →

#	Model	ROC-AUC ↑	Paper
1	ProtoMol Hyb global	0.9140 ±0.0030	ProtoMol: Enhancing Molecular Property Prediction via Prototype-Guided Multimodal Learning
2	MMSG GNN global	0.9080 ±0.0050	ProtoMol: Enhancing Molecular Property Prediction via Prototype-Guided Multimodal Learning
3	MemGNN Hyb ★ proposed	0.9070	This paper
38	Random Forest RF baseline	0.8670 ±0.0080	This paper
43	GIN code ↗ GNN global	0.8646 ±0.0081	Pre-training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information … code ↗
48	Kernel SVM SVM baseline	0.8620	This paper
64	GCN GNN baseline actual #64 vs paper #211 147 places worse	0.8540 ±0.0110 best reported	actual #64 vs paper #211 147 places worse paper 0.7830 ±0.0140 best today: Atom-Motif Contrastive Transformer for Molecular Property Prediction (2023) paper 0.7830 ±0.0140 Δ today 0.0710
75	XGBoost Boost baseline	0.8500 ±0.0080	This paper
93	IRV Trad baseline	0.8380	This paper
109	Bypass Trad baseline	0.8290 ±0.0060	This paper
122	Multitask GNN baseline	0.8240 ±0.0006	This paper
167	Weave GNN baseline	0.8060 ±0.0020	This paper
215	Logistic Regression Trad baseline	0.7810 ±0.0100	This paper

ESOL

Global leaderboard →

#	Model	RMSE ↓	Paper
1	Graphormer-SPIS code ↗ GT global	0.4840 ±0.0050	On Structural Expressive Power of Graph Transformers code ↗
2	Graphormer-SPD code ↗ GT global	0.4920 ±0.0040	On Structural Expressive Power of Graph Transformers code ↗
3	conan-FGW global	0.5140 ±0.0190	Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks
4	MemGNN Hyb baseline	0.5400 ±0.0100	This paper
7	GCN GNN baseline actual #7 vs paper #79 72 places worse	0.5640 ±0.0150 best reported	actual #7 vs paper #79 72 places worse paper 0.9700 ±0.0100 best today: On Structural Expressive Power of Graph Transformers (KDD 2023) paper 0.9700 ±0.0100 Δ today 0.4060
18	GraphSAGE code ↗ GNN global	0.6010 ±0.0080	On Structural Expressive Power of Graph Transformers code ↗
94	Random Forest RF baseline	1.0700 ±0.1900	This paper
106	Multitask GNN baseline	1.1200 ±0.1900	This paper

Lipophilicity

Global leaderboard →

#	Model	RMSE ↓	Paper
1	MemGNN Hyb baseline	0.5560 ±0.0230	This paper
2	ProtoS-L2 Hyb 0.1M global	0.5800 ±0.0160	Optimal Transport Graph Neural Networks
3	ProtoW-L2 Hyb global	0.6040 ±0.0140	Optimal Transport Graph Neural Networks
7	GCN GNN baseline	0.6550 ±0.0360	This paper
9	GIN GNN 0.6M global	0.6580 ±0.0190	Optimal Transport Graph Neural Networks
22	MPNN GNN baseline actual #22 vs paper #24 2 places worse	0.7150 ±0.0350	actual #22 vs paper #24 2 places worse paper 0.7150 ±0.0350 prev best: MoleculeNet (2017) Improvement within reported uncertainty; not highlighted. paper 0.7150 ±0.0350 Δ at pub 0.0040
74	Random Forest RF baseline	0.8760 ±0.0400	This paper

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 at pub — reported value improved over the best known when published Improved at pub — later paper surpassed this result Best today — OK at publication; today's leaderboard best is elsewhere