Heterogeneous graph neural network based on multi-head relation transformation

doi:10.11772/j.issn.1001-9081.2025081065

Abstract

Abstract:

Concerning the problem that the existing Heterogeneous Graph Neural Networks （HGNNs） rely on manually constructed meta-paths and fail to fully capture complex relation semantics， a Multi-Head Relation-Transformed Heterogeneous Graph neural Network （MHRTHGN） method was proposed. In this method， independent relation transformation matrices were constructed for each type of relation， a multi-head attention mechanism was introduced to characterize multi-relation features precisely， and a relation-specific multi-head attention mechanism was designed to fuse multi-relation information at both node-level and relation-level. At the same time， a Feed-Forward Network （FFN） was combined to further enhance semantic expressiveness capability of the model. Experimental results show that on the ACM， IMDB， and DBLP datasets， the proposed method improves the accuracies of node classification and clustering tasks by 3-5 percentage points on average compared with representative baseline models such as HAN （Heterogeneous graph Attention Network） and Graph Convolutional Network （GCN）， validating its effectiveness and performance superiority in multi-relation heterogeneous graph learning. It can be seen that MHRTHGN method can model multi-relation information without relying on meta-path prior， providing a more generalizable solution with higher application potential for heterogeneous graph representation learning.

Key words: Heterogeneous Graph (HG), Graph Neural Network (GNN), graph representation learning, multi-head attention mechanism, relation-awareness

摘要：

针对异构图神经网络（HGNN）在多关系语义建模中依赖人工构造元路径、难以充分捕捉复杂关系语义的问题，提出一种基于多头关系转化的异构图神经网络（MHRTHGN）方法。该方法为不同关系构建独立的关系变换矩阵，引入多头注意力机制精细刻画多关系特征，并设计关系特定多头注意力机制在节点级与关系级两层面融合多关系信息，同时结合前馈网络（FFN）进一步提升表示的语义表达能力。实验结果表明，在ACM、IMDB与DBLP数据集上，与HAN（Heterogeneous graph Attention Network）、图卷积网络（GCN）等代表性模型相比，所提方法在节点分类与节点聚类任务中的准确率平均提升3～5个百分点，验证了模型在多关系异构图学习中的有效性和性能优势。可见，MHRTHGN能在无须元路径先验的条件下有效建模多关系信息，为异构图表示学习提供了一种泛化性更强、应用潜力更大的技术方案。

关键词: 异构图, 图神经网络, 图表示学习, 多头注意力机制, 关系感知

CLC Number:

TP18

Huajuan HUANG, Yongkang FANG, Xiuxi WEI. Heterogeneous graph neural network based on multi-head relation transformation[J]. Journal of Computer Applications, 2026, 46(10): 3083-3092.

黄华娟, 方永康, 韦修喜. 基于多头关系转化的异构图神经网络[J]. 《计算机应用》唯一官方网站, 2026, 46(10): 3083-3092.

Figures/Tables 10

References 29

[1]	Xu H， Yang B， Liu X， et al. Category-aware multi-relation heterogeneous graph neural networks for session-based recommendation［J］. Knowledge-Based Systems， 2022， 251： No.109246.
[2]	Zhao J， Shao M， Tang H， et al. RHGNN： fake reviewer detection based on reinforced heterogeneous graph neural networks［J］. Knowledge-Based Systems， 2023， 280： No.111029.
[3]	Zhang J， Chen S L， Wang Y C. Systematic prediction of synergistic drug combinations through network-based deep learning framework［J］. Expert Systems with Applications， 2025， 270： No.126566.
[4]	马帅，刘建伟，左信. 图神经网络综述［J］. 计算机研究与发展， 2022， 59（1）： 47-80.
	Ma Shuai， Liu Jianwei， Zuo Xin. Survey on graph neural network ［J］. Journal of Computer Research and Development， 2022， 59 （1）： 47-80.
[5]	徐冰冰，岑科廷，黄俊杰，等. 图卷积神经网络综述［J］. 计算机学报， 2020， 43（5）： 755-780.
	Xu Bingbing， Cen Keting， Huang Junjie， et al. A survey on graph convolutional neural network ［J］. Chinese Journal of Computers， 2020， 43 （5）： 755-780.
[6]	Liu J， Ong G P， Chen X. GraphSAGE-based traffic speed forecasting for segment network with sparse data［J］. IEEE Transactions on Intelligent Transportation Systems， 2022， 23（3）： 1755-1766.
[7]	Lai P Y， Dai Q Y， Lu Y H， et al. MIGP： metapath integrated graph prompt neural network［J］. Neural Networks， 2024， 179： No.106595.
[8]	石川，王睿嘉，王啸. 异质信息网络分析与应用综述［J］. 软件学报， 2022， 33（2）： 598-621.
	Shi Chuan， Wang Ruijia， Wang Xiao. Survey on heterogeneous information networks analysis and applications ［J］. Journal of Software， 2022， 33 （2）： 598-621.
[9]	Wei X， Zhang Z， Huang H， et al. An overview on deep clustering［J］. Neurocomputing， 2024， 590： No.127761.
[10]	Huang H， Wang C， Wei X， et al. Deep image clustering： a survey［J］. Neurocomputing， 2024， 599： No.128101.
[11]	Yang X， Yan M， Pan S， et al. Simple and efficient heterogeneous graph neural network［J］. Proceedings of the AAAI Conference on Artificial Intelligence， 2023， 37（9）： 10816-10824.
[12]	Li C， Fu J， Yan Y， et al. Higher order heterogeneous graph neural network based on node attribute enhancement［J］. Expert Systems with Applications， 2024， 238（Pt F）： No.122404.
[13]	Hu Z， Dong Y， Wang K， et al. Heterogeneous graph Transformer［C］// WWW 2020. New York： ACM， 2020： 2704-2710.
[14]	Wang X， Ji H， Shi C， et al. Heterogeneous graph attention network［C］// WWW 2019. New York： ACM， 2019： 2022-2032.
[15]	Wang X， Bo D， Shi C， et al. A survey on heterogeneous graph embedding： methods， techniques， applications and sources［J］. IEEE Transactions on Big Data， 2023， 9（2）： 415-436.
[16]	Gao Y， Zhang P， Zhou C， et al. HGNAS++： efficient architecture search for heterogeneous graph neural networks［J］. IEEE Transactions on Knowledge and Data Engineering， 2023， 35（9）： 9448-9461.
[17]	Wang Z， Yu D， Li Q， et al. SR-HGN： semantic- and relation-aware heterogeneous graph neural network［J］. Expert Systems with Applications， 2023， 224： No.119982.
[18]	Melton J， Krishnan S. muxGNN： multiplex graph neural network for heterogeneous graphs［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2023， 45（9）： 11067-11078.
[19]	Li Z， Li C， Li M， et al. Federated graph Transformer with mixture attentions for secure graph knowledge fusions［J］. Information Fusion， 2025， 118： No.102954.
[20]	Liu Y， Fan L， Wang X， et al. HGBER： heterogeneous graph neural network with bidirectional encoding representation［J］. IEEE Transactions on Neural Networks and Learning Systems， 2024， 35（7）： 9340-9351.
[21]	Zhang C， Song D， Huang C， et al. Heterogeneous graph neural network［C］// KDD 2019. New York： ACM， 2019： 793-803.
[22]	Grover A， Leskovec J. node2vec： scalable feature learning for networks［C］// KDD 2016. New York： ACM， 2016： 855-864.
[23]	Kipf T N， Welling M. Semi-supervised classification with graph convolutional networks［PP/OL］. V4. arXiv （2017-02-22）［2025-01-01］..
[24]	Fu X， Zhang J， Meng Z， et al. MAGNN： metapath aggregated graph neural network for heterogeneous graph embedding［C］// WWW 2020. New York： ACM， 2020： 2331-2341.
[25]	Guan M， Cai X， Shang J， et al. HMSG： heterogeneous graph neural network based on metapath subgraph learning［J］. Knowledge-Based Systems， 2023， 279： No.110930.
[26]	Han M， Zhang H， Li W， et al. Semantic-guided graph neural network for heterogeneous graph embedding［J］. Expert Systems with Applications， 2023， 232： No.120810.
[27]	Han M， Zhang H. KAGNN： graph neural network with kernel alignment for heterogeneous graph learning［J］. Knowledge-Based Systems， 2024， 289： No.111561.
[28]	Zhao Y， Xu S， Duan H. HGNN-BRFE： heterogeneous graph neural network model based on region feature extraction［J］. Electronics， 2024， 13（22）： No.4447.
[29]	Lin W， Chen Z， Chen Y， et al. Heterogeneous graph neural network with adaptive relation reconstruction［J］. Neural Networks， 2025， 187： No.107313.

数据集	模型	Macro-F1				Micro-F1
数据集	模型	δ=20%	δ=40%	δ=60%	δ=80%	δ=20%	δ=40%	δ=60%	δ=80%
ACM	N2V	85.42	86.84	87.45	88.82	85.17	86.46	87.28	88.44
	GCN	90.82	91.26	91.89	91.26	90.67	91.28	91.96	91.26
	HAN	90.43	90.63	91.09	91.25	90.45	90.67	91.07	91.36
	HGNN	90.15	90.42	90.73	91.60	91.23	91.52	91.42	91.66
	MAGNN	88.56	89.38	90.26	91.45	88.35	89.27	90.22	90.46
	HMSG	89.32	90.54	91.09	91.73	89.36	90.37	90.72	91.41
	SGNN	89.88	90.68	91.30	91.76	89.76	90.62	91.20	91.70
	KAGNN	—	—	—	92.19	—	—	—	92.14
	HGNN-BRFE	—	93.58	93.93	94.22	—	93.48	93.84	94.10
	HGNN-AR²	92.42	92.85	94.36	—	92.36	92.93	94.40	—
	MHRTHGN	93.55	94.41	94.96	95.85	93.47	94.23	94.87	95.79
DBLP	N2V	86.70	88.07	88.69	88.93	87.21	88.51	89.09	89.37
	GCN	88.65	89.64	91.17	91.28	89.86	89.99	90.49	91.35
	HAN	92.13	92.64	92.86	91.90	92.60	93.18	93.30	92.50
	HGNN	91.01	91.32	91.53	91.77	91.40	91.79	91.98	92.01
	MAGNN	92.37	93.26	93.06	93.28	93.01	93.70	93.17	93.82
	HMSG	93.04	93.14	93.34	94.05	92.18	93.25	93.86	94.13
	SGNN	93.74	94.09	94.37	94.92	94.23	94.52	94.80	95.30
	KAGNN	—	—	—	94.13	—	—	—	94.60
	HGNN-BRFE	—	94.11	94.21	94.38	—	94.54	94.64	94.78
	HGNN-AR²	92.54	93.06	93.73	—	93.02	93.44	94.15	—
	MHRTHGN	92.61	92.88	94.13	95.80	93.31	93.49	94.66	96.07
IMDB	N2V	49.00	50.63	51.65	51.49	49.94	51.77	52.79	52.72
	GCN	60.34	61.06	61.60	61.62	60.55	60.72	61.86	61.76
	HAN	58.33	57.69	62.75	63.79	60.33	60.53	62.96	64.23
	HGNN	60.02	60.35	61.30	62.52	61.73	62.58	62.95	63.64
	MAGNN	57.28	58.37	58.52	58.29	57.40	58.56	58.68	58.38
	HMSG	59.16	59.87	60.12	61.53	59.89	60.56	60.85	61.34
	SGNN	59.87	60.62	60.97	61.73	60.25	61.03	61.16	61.98
	KAGNN	—	—	—	60.60	—	—	—	60.71
	HGNN-BRFE	—	62.41	62.60	62.59	—	62.47	62.68	52.69
	HGNN-AR²	52.51	53.23	54.12	—	61.36	63.04	64.13	—
	MHRTHGN	60.44	61.99	64.66	65.55	60.48	61.93	64.95	66.90

数据集	模型	Macro-F1				Micro-F1
数据集	模型	δ=20%	δ=40%	δ=60%	δ=80%	δ=20%	δ=40%	δ=60%	δ=80%
ACM	N2V	85.42	86.84	87.45	88.82	85.17	86.46	87.28	88.44
	GCN	90.82	91.26	91.89	91.26	90.67	91.28	91.96	91.26
	HAN	90.43	90.63	91.09	91.25	90.45	90.67	91.07	91.36
	HGNN	90.15	90.42	90.73	91.60	91.23	91.52	91.42	91.66
	MAGNN	88.56	89.38	90.26	91.45	88.35	89.27	90.22	90.46
	HMSG	89.32	90.54	91.09	91.73	89.36	90.37	90.72	91.41
	SGNN	89.88	90.68	91.30	91.76	89.76	90.62	91.20	91.70
	KAGNN	—	—	—	92.19	—	—	—	92.14
	HGNN-BRFE	—	93.58	93.93	94.22	—	93.48	93.84	94.10
	HGNN-AR²	92.42	92.85	94.36	—	92.36	92.93	94.40	—
	MHRTHGN	93.55	94.41	94.96	95.85	93.47	94.23	94.87	95.79
DBLP	N2V	86.70	88.07	88.69	88.93	87.21	88.51	89.09	89.37
	GCN	88.65	89.64	91.17	91.28	89.86	89.99	90.49	91.35
	HAN	92.13	92.64	92.86	91.90	92.60	93.18	93.30	92.50
	HGNN	91.01	91.32	91.53	91.77	91.40	91.79	91.98	92.01
	MAGNN	92.37	93.26	93.06	93.28	93.01	93.70	93.17	93.82
	HMSG	93.04	93.14	93.34	94.05	92.18	93.25	93.86	94.13
	SGNN	93.74	94.09	94.37	94.92	94.23	94.52	94.80	95.30
	KAGNN	—	—	—	94.13	—	—	—	94.60
	HGNN-BRFE	—	94.11	94.21	94.38	—	94.54	94.64	94.78
	HGNN-AR²	92.54	93.06	93.73	—	93.02	93.44	94.15	—
	MHRTHGN	92.61	92.88	94.13	95.80	93.31	93.49	94.66	96.07
IMDB	N2V	49.00	50.63	51.65	51.49	49.94	51.77	52.79	52.72
	GCN	60.34	61.06	61.60	61.62	60.55	60.72	61.86	61.76
	HAN	58.33	57.69	62.75	63.79	60.33	60.53	62.96	64.23
	HGNN	60.02	60.35	61.30	62.52	61.73	62.58	62.95	63.64
	MAGNN	57.28	58.37	58.52	58.29	57.40	58.56	58.68	58.38
	HMSG	59.16	59.87	60.12	61.53	59.89	60.56	60.85	61.34
	SGNN	59.87	60.62	60.97	61.73	60.25	61.03	61.16	61.98
	KAGNN	—	—	—	60.60	—	—	—	60.71
	HGNN-BRFE	—	62.41	62.60	62.59	—	62.47	62.68	52.69
	HGNN-AR²	52.51	53.23	54.12	—	61.36	63.04	64.13	—
	MHRTHGN	60.44	61.99	64.66	65.55	60.48	61.93	64.95	66.90

模型	ACM		DBLP		IMDB
模型	ARI	NMI	ARI	NMI	ARI	NMI
N2V	67.53	69.86	76.52	71.02	16.02	20.22
GCN	20.82	34.07	76.67	67.91	17.23	17.86
HAN	73.63	68.92	83.13	77.27	21.53	20.32
HGNN	70.79	72.74	78.83	80.27	18.88	19.64
HMSG	70.51	66.16	80.52	76.36	14.77	14.52
MAGNN	69.18	66.09	83.19	77.55	13.78	14.13
SGNN	71.16	67.95	83.69	77.81	16.29	15.45
KAGNN	71.63	67.50	82.19	77.20	14.49	14.89
HGNN-BRFE	78.92	74.11	85.58	79.95	13.77	13.21
HGNN-AR²	—	—	—	—	—	—
MHRTHGN	76.07	74.40	85.61	80.04	28.83	24.70

模型	ACM		DBLP		IMDB
模型	ARI	NMI	ARI	NMI	ARI	NMI
N2V	67.53	69.86	76.52	71.02	16.02	20.22
GCN	20.82	34.07	76.67	67.91	17.23	17.86
HAN	73.63	68.92	83.13	77.27	21.53	20.32
HGNN	70.79	72.74	78.83	80.27	18.88	19.64
HMSG	70.51	66.16	80.52	76.36	14.77	14.52
MAGNN	69.18	66.09	83.19	77.55	13.78	14.13
SGNN	71.16	67.95	83.69	77.81	16.29	15.45
KAGNN	71.63	67.50	82.19	77.20	14.49	14.89
HGNN-BRFE	78.92	74.11	85.58	79.95	13.77	13.21
HGNN-AR²	—	—	—	—	—	—
MHRTHGN	76.07	74.40	85.61	80.04	28.83	24.70

变体	ACM				DBLP				IMDB
变体	Ma-F1	Mi-F1	ARI	NMI	Ma-F1	Mi-F1	ARI	NMI	Ma-F1	Mi-F1	ARI	NMI
MHRTHGN	93.55	93.47	76.07	74.40	92.61	93.31	85.61	80.04	60.44	60.48	28.83	24.70
w/o F	92.92	92.81	75.60	74.21	91.88	92.52	85.40	79.05	60.05	60.27	26.89	23.27
w/o R	92.50	92.39	75.48	74.24	91.77	92.35	84.83	79.37	60.13	60.32	25.46	23.40
w/o M	92.37	92.22	75.49	74.20	91.55	93.14	84.78	79.34	60.23	60.26	28.21	24.50