Directed Homophily-Aware Graph Neural Network

Aihu Zhang, Jiaxing Xu, Mengcheng Lan, Shili Xiang, Yiping Ke

Abstract

Graph Neural Networks (GNNs) have achieved significant success in various learning tasks on graph-structured data. Nevertheless, most GNNs struggle to generalize to heterophilic neighborhoods. Additionally, many GNNs ignore the directional nature of real-world graphs, resulting in suboptimal performance on directed graphs with asymmetric structures. In this work, we propose Directed Homophily-aware Graph Neural Network (DHGNN), a novel framework that addresses these limitations by incorporating homophily-aware and direction-sensitive components. DHGNN employs a resettable gating mechanism to adaptively modulate message contributions based on homophily levels and informativeness, and a structure-aware noise-tolerant fusion module to effectively integrate node representations from the original and reverse directions. Extensive experiments on both homophilic and heterophilic directed graph datasets demonstrate that DHGNN outperforms state-of-the-art methods in node classification and link prediction. In particular, DHGNN improves over the best baseline by up to 15.07\

Benchmark Results

Tested on 3 datasets 17 baseline values reported too low

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

Heterophilic Graphs

Chameleon

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	Hetero-S (GAT 8-layer) code ↗ Hyb global	0.8693	The Heterophilic Snowflake Hypothesis: Training and Empowering GNNs for Heterophilic Graphs code ↗
2	Dir-GNN+CNA code ↗ GNN global	0.8586 ±0.0180	Graph Neural Networks Need Cluster-Normalize-Activate Modules code ↗
3	GAT GNN baseline actual #3 vs paper #223 220 places worse	0.8560 best reported	actual #3 vs paper #223 220 places worse paper 0.4556 ±0.0316 best at pub: The Heterophilic Snowflake Hypothesis: Training and Empowering GNNs for Heterophilic Graphs (KDD 2024) paper 0.4556 ±0.0316 Δ at pub 0.4004
7	Dir-GNN GNN baseline actual #7 vs paper #15 8 places worse	0.8048 ±0.0146 best reported	actual #7 vs paper #15 8 places worse paper 0.7971 ±0.0126 best at pub: Graph Neural Networks Need Cluster-Normalize-Activate Modules (NeurIPS 2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7971 ±0.0126 Δ at pub 0.0077
10	DHGNN GNN ★ proposed	0.8011 ±0.0173	This paper
14	NDDGNN GNN baseline	0.7979 ±0.0104	This paper
25	FSGNN GNN baseline	0.7827 ±0.0128	This paper
28	ACM-GCN GNN baseline actual #28 vs paper #28	0.7476 ±0.0220	actual #28 vs paper #28 paper 0.7476 ±0.0220 prev best: Diffusion on Graph: Augmentation of Graph Structure for Node Classification (TMLR 2025) paper 0.7476 ±0.0220 Δ at pub 0.0576
44	G^2 GNN baseline	0.7140 ±0.0238	This paper
45	GloGNN GNN baseline actual #45 vs paper #147 102 places worse	0.7133 ±0.0242 best reported	actual #45 vs paper #147 102 places worse paper 0.5788 ±0.0176 best at pub: GPNet: Simplifying Graph Neural Networks via Multi-channel Geometric Polynomials (2022) best today: $P^2$GNN: Two Prototype Sets to boost GNN Performance (2026) paper 0.5788 ±0.0176 Δ at pub 0.1333 Δ today 0.1345
52	GCN GNN baseline actual #52 vs paper #95 43 places worse	0.7033 best reported	actual #52 vs paper #95 43 places worse paper 0.6482 ±0.0224 best at pub: Understanding Heterophily for Graph Neural Networks (ICML 2024) paper 0.6482 ±0.0224 Δ at pub 0.0551
64	LINKX GNN baseline	0.6842 ±0.0132	This paper
72	GPRGNN GNN baseline actual #72 vs paper #110 38 places worse	0.6748 ±0.0040 best reported	actual #72 vs paper #110 38 places worse paper 0.6285 ±0.0290 best at pub: LON-GNN: Spectral GNNs with Learnable Orthonormal Basis (2023) best today: $P^2$GNN: Two Prototype Sets to boost GNN Performance (2026) paper 0.6285 ±0.0290 Δ at pub 0.0443 Δ today 0.0463
108	SAGE code ↗ GNN global	0.6340 ±0.0300	Scale-aware Message Passing For Graph Node Classification code ↗
133	H2GCN GNN baseline actual #133 vs paper #138 5 places worse	0.6030 ±0.0231 best reported	actual #133 vs paper #138 5 places worse paper 0.5939 ±0.0098 best at pub: Optimization-Induced Graph Implicit Nonlinear Diffusion (ICML 2022) Difference within reported uncertainty; not counted as underperformance. paper 0.5939 ±0.0098 Δ at pub 0.0091
147	MagNet GNN baseline actual #147 vs paper #147	0.5832 ±0.0287	actual #147 vs paper #147 paper 0.5832 ±0.0287 prev best: Scale-aware Message Passing For Graph Node Classification (2024) Improvement within reported uncertainty; not highlighted. paper 0.5832 ±0.0287 Δ at pub 0.0012
169	MLP Trad baseline actual #169 vs paper #216 47 places worse	0.5429 best reported	actual #169 vs paper #216 47 places worse paper 0.4636 ±0.0257 best at pub: Understanding Heterophily for Graph Neural Networks (ICML 2024) paper 0.4636 ±0.0257 Δ at pub 0.0793
179	DI-GCN GNN baseline	0.5224 ±0.0365	This paper

Squirrel

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	IBG-NN GNN global	0.7763 ±0.0179	Efficient Learning on Large Graphs using a Densifying Regularity Lemma
2	Dir-GNN+CNA code ↗ GNN global	0.7747 ±0.0128	Graph Neural Networks Need Cluster-Normalize-Activate Modules code ↗
3	DHGNN GNN ★ proposed	0.7684 ±0.0161	This paper
7	Dir-GNN GNN baseline actual #7 vs paper #8 1 place worse	0.7560 ±0.0190 best reported	actual #7 vs paper #8 1 place worse paper 0.7531 ±0.0192 best at pub: Scale-aware Message Passing For Graph Node Classification (2024) Difference within reported uncertainty; not counted as underperformance. paper 0.7531 ±0.0192 Δ at pub 0.0029
8	NDDGNN GNN baseline	0.7538 ±0.0195	This paper
11	FSGNN GNN baseline actual #11 vs paper #11	0.7416 best reported	actual #11 vs paper #11 paper 0.7410 ±0.0189 best at pub: Simplifying approach to Node Classification in Graph Neural Networks (2021) Difference within reported uncertainty; not counted as underperformance. paper 0.7410 ±0.0189 Δ at pub 0.0006
19	GloGNN GNN baseline actual #19 vs paper #19	0.7121 ±0.0184	actual #19 vs paper #19 paper 0.7121 ±0.0184 prev best: GPNet: Simplifying Graph Neural Networks via Multi-channel Geometric Polynomials (2022) paper 0.7121 ±0.0184 Δ at pub 0.1333
25	ACM-GCN GNN baseline actual #25 vs paper #25	0.6740 ±0.0221	actual #25 vs paper #25 paper 0.6740 ±0.0221 prev best: From Node Interaction to Hop Interaction: New Effective and Scalable Graph Learning Paradigm (2022) paper 0.6740 ±0.0221 Δ at pub 0.1300
31	G^2 GNN baseline	0.6426 ±0.0238	This paper
39	H2GCN GNN baseline actual #39 vs paper #184 145 places worse	0.6190 ±0.0140 best reported	actual #39 vs paper #184 145 places worse paper 0.3790 ±0.0002 best at pub: Efficient Learning on Large Graphs using a Densifying Regularity Lemma (2025) paper 0.3790 ±0.0002 Δ at pub 0.2400
40	LINKX GNN baseline	0.6181 ±0.0180	This paper
46	GCN GNN baseline actual #46 vs paper #77 31 places worse	0.6067 best reported	actual #46 vs paper #77 31 places worse paper 0.5343 ±0.0201 best at pub: Understanding Heterophily for Graph Neural Networks (ICML 2024) paper 0.5343 ±0.0201 Δ at pub 0.0724
67	GAT GNN baseline actual #67 vs paper #173 106 places worse	0.5630 ±0.0220 best reported	actual #67 vs paper #173 106 places worse paper 0.3914 ±0.0288 best at pub: Restructuring Graphs for Higher Homophily via Adaptive Spectral Clustering (AAAI 2022) paper 0.3914 ±0.0288 Δ at pub 0.1716
76	GPRGNN GNN baseline actual #76 vs paper #76	0.5435 ±0.0087	actual #76 vs paper #76 paper 0.5435 ±0.0087 prev best: LON-GNN: Spectral GNNs with Learnable Orthonormal Basis (2023) paper 0.5435 ±0.0087 Δ at pub 0.0420
120	SAGE code ↗ GNN global	0.4460 ±0.0130	Scale-aware Message Passing For Graph Node Classification code ↗
136	MagNet GNN baseline actual #136 vs paper #173 37 places worse	0.4270 ±0.0150 best reported	actual #136 vs paper #173 37 places worse paper 0.3901 ±0.0193 best at pub: Breaking the Entanglement of Homophily and Heterophily in Semi-supervised Node Classification (2023) paper 0.3901 ±0.0193 Δ at pub 0.0369
172	MLP Trad baseline actual #172 vs paper #228 56 places worse	0.3930 ±0.0079 best reported	actual #172 vs paper #228 56 places worse paper 0.3414 ±0.0194 best at pub: Understanding Heterophily for Graph Neural Networks (ICML 2024) best today: Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification (2026) paper 0.3414 ±0.0194 Δ at pub 0.0249 Δ today 0.0516
189	DI-GCN GNN baseline	0.3774 ±0.0154	This paper

Heterophily Benchmark

Roman-empire

Global leaderboard →

#	Model	Accuracy ↑	Paper
1	Dir-Poly code ↗ GNN global	0.9451 ±0.0022	Graph Transductive Sharpening: Leveraging Unlabeled Predictions in Node Classification code ↗
2	D-CDE code ↗ GNN global	0.9387 ±0.0041	Distributed-Order Fractional Graph Operating Network code ↗
3	LargeScaleNet code ↗ Hyb global	0.9358 ±0.0024	Scale-aware Message Passing For Graph Node Classification code ↗
14	NDDGNN GNN baseline	0.9176 ±0.0027	This paper
24	Dir-GNN GNN baseline	0.9123 ±0.0032	This paper
26	GraphSAGE code ↗ GNN global	0.9106 ±0.0027	Directed Semi-Simplicial Learning with Applications to Brain Activity Decoding code ↗
36	DHGNN GNN ★ proposed	0.8905 ±0.0053	This paper
47	MagNet GNN baseline	0.8807 ±0.0027	This paper
77	GCN GNN baseline actual #77 vs paper #113 36 places worse	0.8272 ±0.0082 best reported	actual #77 vs paper #113 36 places worse paper 0.7369 ±0.0074 best at pub: HyperAggregation: Aggregating over Graph Edges with Hypernetworks (2024) paper 0.7369 ±0.0074 Δ at pub 0.0903
80	G^2 GNN baseline	0.8216 ±0.0078	This paper
83	GAT GNN baseline actual #83 vs paper #121 38 places worse	0.8145 ±0.0009 best reported	actual #83 vs paper #121 38 places worse paper 0.7116 ±0.0063 best at pub: Graph Neural Networks for Graphs with Heterophily: A Survey (2022) best today: Gumbel-MPNN: Graph Rewiring with Gumbel-Softmax (2025) paper 0.7116 ±0.0063 Δ at pub 0.0971 Δ today 0.1029
92	FSGNN GNN baseline	0.7992 ±0.0056	This paper
115	ACM-GCN GNN baseline actual #115 vs paper #124 9 places worse	0.7189 ±0.0061 best reported	actual #115 vs paper #124 9 places worse paper 0.6966 ±0.0062 best at pub: Graph Neural Networks for Graphs with Heterophily: A Survey (2022) best today: ATLAS: Adaptive Topology-based Learning at Scale for Homophilic and Heterophilic Graphs (2025) paper 0.6966 ±0.0062 Δ at pub 0.0176 Δ today 0.0223
129	H2GCN GNN baseline actual #129 vs paper #164 35 places worse	0.6809 ±0.0029 best reported	actual #129 vs paper #164 35 places worse paper 0.6011 ±0.0052 best at pub: Unleashing the Potential of Fractional Calculus in Graph Neural Networks with FROND (ICLR 2024) best today: Distributed-Order Fractional Graph Operating Network (NeurIPS 2024) paper 0.6011 ±0.0052 Δ at pub 0.0798 Δ today 0.0798
133	GPRGNN GNN baseline actual #133 vs paper #144 11 places worse	0.6704 ±0.0092 best reported	actual #133 vs paper #144 11 places worse paper 0.6485 ±0.0027 best at pub: NTFormer: A Composite Node Tokenized Graph Transformer for Node Classification (2024) paper 0.6485 ±0.0027 Δ at pub 0.0219
135	MLP Trad baseline actual #135 vs paper #141 6 places worse	0.6665 ±0.0047 best reported	actual #135 vs paper #141 6 places worse paper 0.6576 ±0.0042 best at pub: HyperAggregation: Aggregating over Graph Edges with Hypernetworks (2024) paper 0.6576 ±0.0042 Δ at pub 0.0089
154	GloGNN GNN baseline actual #154 vs paper #167 13 places worse	0.6385 ±0.0049 best reported	actual #154 vs paper #167 13 places worse paper 0.5963 ±0.0069 best at pub: Unleashing the Potential of Fractional Calculus in Graph Neural Networks with FROND (ICLR 2024) best today: Distributed-Order Fractional Graph Operating Network (NeurIPS 2024) paper 0.5963 ±0.0069 Δ at pub 0.0422 Δ today 0.0422
168	LINKX GNN baseline actual #168 vs paper #202 34 places worse	0.5914 ±0.0045 best reported	actual #168 vs paper #202 34 places worse paper 0.3755 ±0.0036 best at pub: Graph Neural Networks for Graphs with Heterophily: A Survey (2022) paper 0.3755 ±0.0036 Δ at pub 0.2159
183	DI-GCN GNN baseline	0.5271 ±0.0032	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 Best today — OK at publication; today's leaderboard best is elsewhere