Graph Representation Learning via Aggregation Enhancement

Maxim Fishman, Chaim Baskin, Evgenii Zheltonozhskii, Almog David, Ron Banner, A. Mendelson

Abstract

Graph neural networks (GNNs) have become a powerful tool for processing graph-structured data but still face challenges in effectively aggregating and propagating information between layers, which limits their performance. We tackle this problem with the kernel regression (KR) approach, using KR loss as the primary loss in self-supervised settings or as a regularization term in supervised settings. We show substantial performance improvements compared to state-of-the-art in both scenarios on multiple transductive and inductive node classification datasets, especially for deep networks. As opposed to mutual information (MI), KR loss is convex and easy to estimate in high-dimensional cases, even though it indirectly maximizes the MI between its inputs. Our work highlights the potential of KR to advance the field of graph representation learning and enhance the performance of GNNs. The code to reproduce our experiments is available at https://github.com/Anonymous1252022/KR_for_GNNs

Benchmark Results

Tested on 10 datasets Non-standard metric on 3 datasets

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

Classic

Amazon-Photo

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	GraphTARIF code ↗ Hyb global	0.9703 ±0.0019	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
2	GAT code ↗ GNN global	0.9669 ±0.0014	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
3	GAT+T code ↗ Hyb global	0.9660 ±0.0007	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
7	SAGE code ↗ GNN global	0.9643 ±0.0027	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
28	GIRL GNN ★ proposed	0.9560 ±0.0010	This paper
172	BGRL GNN baseline actual #172 vs paper #190 18 places worse	0.9337 ±0.0021 best reported	actual #172 vs paper #190 18 places worse paper 0.9320 ±0.0030 best at pub: Learning Graph Augmentations to Learn Graph Representations (2022) best today: Wiener Graph Deconvolutional Network Improves Graph Self-Supervised Learning (AAAI 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.9320 ±0.0030 Δ today 0.0017
203	GRACE GNN baseline actual #203 vs paper #324 121 places worse	0.9313 ±0.0017 best reported	actual #203 vs paper #324 121 places worse paper 0.9220 ±0.0020 best at pub: Learning Graph Augmentations to Learn Graph Representations (2022) best today: Learn Beneficial Noise as Graph Augmentation (ICML 2025) paper 0.9220 ±0.0020 Δ at pub 0.0060 Δ today 0.0093
213	DGI GNN baseline actual #213 vs paper #366 153 places worse	0.9309 ±0.0008 best reported	actual #213 vs paper #366 153 places worse paper 0.9160 ±0.0020 best at pub: Graph Contrastive Learning with Adaptive Augmentation (WWW 2020) best today: Noise-Resilient Unsupervised Graph Representation Learning via Multi-Hop Feature Quality Estimation (2024) paper 0.9160 ±0.0020 Δ at pub 0.0001 Δ today 0.0149
281	BGCL GNN baseline	0.9250 ±0.0020	This paper
390	MLP code ↗ Trad global	0.9128 ±0.0063	Clarify Confused Nodes via Separated Learning code ↗

Amazon-Computers

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	GraphTARIF code ↗ Hyb global	0.9461 ±0.0017	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
2	GCN code ↗ GNN global	0.9412 ±0.0008	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
3	GAT code ↗ GNN global	0.9398 ±0.0022	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
8	SAGE code ↗ GNN global	0.9351 ±0.0006	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
31	GIRL GNN ★ proposed	0.9180	This paper
79	BGRL GNN baseline actual #79 vs paper #81 2 places worse	0.9030 ±0.0020	actual #79 vs paper #81 2 places worse paper 0.9030 ±0.0020 prev best: Augmentation-Free Self-Supervised Learning on Graphs (AAAI 2021) paper 0.9030 ±0.0020 Δ at pub 0.0061
139	GRACE GNN baseline actual #139 vs paper #139	0.8950 ±0.0040	actual #139 vs paper #139 paper 0.8950 ±0.0040 prev best: Barlow Graph Auto-Encoder for Unsupervised Network Embedding (2021) paper 0.8950 ±0.0040 Δ at pub 0.0204
210	DGI GNN baseline actual #210 vs paper #350 140 places worse	0.8400 ±0.0050	actual #210 vs paper #350 140 places worse paper 0.8400 ±0.0050 prev best: Graph Contrastive Learning with Adaptive Augmentation (WWW 2020) best today: Bootstrap Latents of Nodes and Neighbors for Graph Self-Supervised Learning (2024) Improvement within reported uncertainty; not highlighted. paper 0.8400 ±0.0050 Δ at pub 0.0005 Δ today 0.0398
320	MLP code ↗ Trad global	0.8463	GraphFM: A generalist graph transformer that learns transferable representations across diverse … code ↗

Global leaderboard →

Non-default metric. This paper reports Accuracy; the dataset default is F1. Values below use Accuracy only.

Sorted by Accuracy. Highlighted rows are from other papers.

#	Model	Accuracy ↑	Paper
1	GAP-EDP code ↗ GNN global	0.9870 ±0.0003	GAP: Differentially Private Graph Neural Networks with Aggregation Perturbation code ↗
2	NE-ASGCN ext data Hyb global	0.9758 ±0.0027	Neighbor Enhanced Graph Convolutional Networks for Node Classification and Recommendation
3	HSGT GT global	0.9730 ±0.0024	Hierarchical Transformer for Scalable Graph Learning
21	GraphSAGE GNN global	0.9653 ±0.0011	Hierarchical Transformer for Scalable Graph Learning
42	GraphMAE GNN baseline actual #42 vs paper #42	0.9601 ±0.0008 best reported	actual #42 vs paper #42 paper 0.9600 ±0.0010 best today: Generative and Contrastive Paradigms Are Complementary for Graph Self-Supervised Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.9600 ±0.0010 Δ today 0.0001
74	GRACE GNN baseline actual #74 vs paper #105 31 places worse	0.9472 ±0.0004 best reported	actual #74 vs paper #105 31 places worse paper 0.9170 best today: Generative and Contrastive Paradigms Are Complementary for Graph Self-Supervised Learning (2023) paper 0.9170 Δ today 0.0302
77	DGI GNN baseline actual #77 vs paper #77	0.9403 ±0.0010 best reported	actual #77 vs paper #77 paper 0.9400 ±0.0010 best today: Generative and Contrastive Paradigms Are Complementary for Graph Self-Supervised Learning (2023) Difference within reported uncertainty; not counted as underperformance. paper 0.9400 ±0.0010 Δ today 0.0003
93	KR GNN baseline	0.9310 ±0.0010	This paper
102	GIRL GNN ★ proposed	0.9210	This paper
147	MLP code ↗ Trad global	0.8240 ±0.0010	GAP: Differentially Private Graph Neural Networks with Aggregation Perturbation code ↗

Coauthor-Physics

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	GrokFormer code ↗ GT global	0.9831 ±0.0018	GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers code ↗
2	PolyFormer code ↗ GT global	0.9808 ±0.0027	GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers code ↗
3	SGFormer code ↗ GT global	0.9796 ±0.0081	GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers code ↗
9	GCN code ↗ GNN global	0.9774 ±0.0035	GrokFormer: Graph Fourier Kolmogorov-Arnold Transformers code ↗
39	GIRL GNN ★ proposed	0.9660	This paper
58	MLP code ↗ Trad global	0.9597 ±0.0007	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
69	BGRL GNN baseline actual #69 vs paper #84 15 places worse	0.9580 ±0.0040 best reported	actual #69 vs paper #84 15 places worse paper 0.9570 ±0.0010 best at pub: Learning Graph Augmentations to Learn Graph Representations (2022) best today: Unifying Graph Contrastive Learning with Flexible Contextual Scopes (2022) Difference within reported uncertainty; not counted as underperformance. paper 0.9570 ±0.0010 Δ today 0.0010
167	DGI GNN baseline actual #167 vs paper #174 7 places worse	0.9489 ±0.0009 best reported	actual #167 vs paper #174 7 places worse paper 0.9450 ±0.0050 best at pub: Graph Contrastive Learning with Adaptive Augmentation (WWW 2020) best today: Balancing Graph Embedding Smoothness in Self-Supervised Learning via Information-Theoretic Decomposition (WWW 2025) Difference within reported uncertainty; not counted as underperformance. paper 0.9450 ±0.0050 Δ at pub 0.0001 Δ today 0.0039

PPI

Global leaderboard →

Non-default metric. This paper reports Accuracy; the dataset default is F1. Values below use Accuracy only.

Sorted by Accuracy. Highlighted rows are from other papers.

#	Model	Accuracy ↑	Paper
1	Cluster-GCN code ↗ GNN global	0.9936	Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks code ↗
2	GaAN code ↗ GNN global	0.9871	Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks code ↗
3	GeniePath code ↗ GNN global	0.9850	Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks code ↗
7	GraphSAGE code ↗ GNN global	0.9580 ±0.0010	Weisfeiler and Leman Go Infinite: Spectral and Combinatorial Pre-Colorings code ↗
12	GIRL GNN ★ proposed	0.8790 ±0.0010	This paper
13	KR GNN baseline	0.8680 ±0.0010	This paper
15	GraphMAE GNN baseline	0.7450 ±0.0030	This paper
17	BGRL GNN baseline	0.7050 ±0.0010	This paper
18	SimGRACE GNN baseline	0.7030 ±0.0120	This paper
20	GRACE GNN baseline	0.6620 ±0.0010	This paper
24	DGI GNN baseline	0.6380 ±0.0020	This paper

WikiCS

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	GCN GNN global	0.9302 ±0.0011	Augmentation-Free Graph Contrastive Learning with Performance Guarantee
2	RSE + RGCN (70b) ext data global	0.8852 ±0.0019	Unleashing the Potential of Text-attributed Graphs: Automatic Relation Decomposition via Large …
3	LEMP+T ext data Hyb global	0.8825	Language Models as Messengers: Enhancing Message Passing in Heterophilic Graph Learning
9	SAGE GNN global	0.8771	Language Models as Messengers: Enhancing Message Passing in Heterophilic Graph Learning
50	GIRL GNN ★ proposed	0.8330	This paper
94	GRACE GNN baseline actual #94 vs paper #96 2 places worse	0.8014 ±0.0048 best reported	actual #94 vs paper #96 2 places worse paper 0.8010 ±0.0050 best at pub: Large-Scale Representation Learning on Graphs via Bootstrapping (ICLR 2021) best today: Adaptive Negative Scheduling for Graph Contrastive Learning (2026) Difference within reported uncertainty; not counted as underperformance. paper 0.8010 ±0.0050 Δ at pub 0.0004 Δ today 0.0004
100	BGRL GNN baseline	0.8000 ±0.0010	This paper
151	DGI GNN baseline actual #151 vs paper #226 75 places worse	0.7865 ±0.0690 best reported	actual #151 vs paper #226 75 places worse paper 0.7540 ±0.0010 best at pub: Learning Graph Augmentations to Learn Graph Representations (2022) best today: Quantizing Text-attributed Graphs for Semantic-Structural Integration (KDD 2025) Difference within reported uncertainty; not counted as underperformance. paper 0.7540 ±0.0010 Δ today 0.0325
237	MLP Trad global	0.7469 ±0.0085	View Space: Learning Representation across Arbitrary Graphs

OGB

ogbn-products

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	Jacobi GNN global	0.8968	Language Models as Messengers: Enhancing Message Passing in Heterophilic Graph Learning
2	GPR GNN global	0.8953	Language Models as Messengers: Enhancing Message Passing in Heterophilic Graph Learning
3	LEMP4HG ext data Hyb global	0.8939	Language Models as Messengers: Enhancing Message Passing in Heterophilic Graph Learning
4	SAGE GNN global	0.8933	Language Models as Messengers: Enhancing Message Passing in Heterophilic Graph Learning
239	+FA GNN baseline	0.7630 ±0.0010	This paper
247	KR GNN baseline	0.7580 ±0.0010	This paper
248	None GNN baseline	0.7580 ±0.0030	This paper
255	WD GNN baseline	0.7560 ±0.0030	This paper
317	GRACE GNN baseline actual #317 vs paper #344 27 places worse	0.7210 ±0.0070 best reported	actual #317 vs paper #344 27 places worse paper 0.6810 ±0.0010 best today: Smoothed Graph Contrastive Learning via Seamless Proximity Integration (2024) paper 0.6810 ±0.0010 Δ today 0.0400
325	GIRL GNN ★ proposed	0.7150 ±0.0030	This paper
351	MLP Trad global	0.6605 ±0.0020	Latent Graph Inference using Product Manifolds

Heterophilic Graphs

Texas

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	P^2GNN Hyb global	0.9672 ±0.0147	$P^2$GNN: Two Prototype Sets to boost GNN Performance
2	DoG code ↗ ext data GNN global	0.9660 ±0.0210	Diffusion on Graph: Augmentation of Graph Structure for Node Classification code ↗
3	GPF-plus code ↗ ext data GNN global	0.9583 ±0.0419	ProG: A Graph Prompt Learning Benchmark code ↗
34	MLP Trad global	0.9230 ±0.0070	Design your own universe: a physics-informed agnostic method for enhancing graph neural networks
53	GraphSAGE code ↗ GNN global	0.9000 ±0.0365	Graph Neural Networks Need Cluster-Normalize-Activate Modules code ↗
279	None GNN baseline	0.7780 ±0.0120	This paper
284	KR GNN baseline	0.7730 ±0.0230	This paper
301	WD GNN baseline	0.7620 ±0.0230	This paper
304	DGI GNN baseline actual #304 vs paper #574 270 places worse	0.7581 ±0.0503 best reported	actual #304 vs paper #574 270 places worse paper 0.5410 best today: Adaptive Graph Mixture of Residual Experts: Unsupervised Learning on Diverse Graphs with Heterogeneous Specialization (2025) paper 0.5410 Δ today 0.2171
307	GIRL GNN ★ proposed	0.7570	This paper
352	+FA GNN baseline	0.7300 ±0.0330	This paper
448	SELENE GNN baseline	0.6430	This paper

Actor

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	DiGGR Hyb global	0.4535 ±0.0353	Disentangled Generative Graph Representation Learning
2	TRIGON code ↗ Hyb global	0.4402 ±0.0028	Dynamic Triangulation-Based Graph Rewiring for Graph Neural Networks code ↗
3	DoG code ↗ ext data GNN global	0.4320 ±0.0120	Diffusion on Graph: Augmentation of Graph Structure for Node Classification code ↗
42	MLP Trad global	0.4026 ±0.0108	Heterophily-Aware Graph Attention Network
68	GraphSAGE code ↗ GNN global	0.3899 ±0.0085	A Quasi-Wasserstein Loss for Learning Graph Neural Networks code ↗
232	KR GNN baseline	0.3510 ±0.0030	This paper
259	GIRL ext data GNN ★ proposed	0.3420	This paper
267	SELENE ext data GNN baseline	0.3410	This paper
335	DGI ext data GNN baseline actual #335 vs paper #409 74 places worse	0.3067 ±0.0063 best reported	actual #335 vs paper #409 74 places worse paper 0.2720 best today: Toward General Digraph Contrastive Learning: A Dual Spatial Perspective (2025) paper 0.2720 Δ today 0.0347

Heterogeneous Graph Benchmarks

DBLP

Global leaderboard →

Non-default metric. This paper reports Accuracy; the dataset default is Micro-F1. Values below use Accuracy only.

Sorted by Accuracy. Highlighted rows are from other papers.

#	Model	Accuracy ↑	Paper
1	HG-SCM Hyb global	0.9490 ±0.0015	Towards Human-like Perception: Learning Structural Causal Model in Heterogeneous Graph
2	SeHGNN GNN global	0.9489 ±0.0018	Towards Human-like Perception: Learning Structural Causal Model in Heterogeneous Graph
3	AdvSynGNN Hyb global	0.9486 ±0.0012	AdvSynGNN: Structure-Adaptive Graph Neural Nets via Adversarial Synthesis and Self-Corrective …
36	GCN code ↗ GNN global	0.8758 ±0.0020	I'm Me, We're Us, and I'm Us: Tri-directional Contrastive Learning on Hypergraphs code ↗
41	GIRL GNN ★ proposed	0.8590 ±0.0010	This paper
54	GRACE GNN baseline	0.8420 ±0.0010	This paper
65	MLP code ↗ Trad global	0.8118 ±0.0020	I'm Me, We're Us, and I'm Us: Tri-directional Contrastive Learning on Hypergraphs code ↗

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