基于属性补全的异质图表示学习算法

doi:10.12068/j.issn.1005-3026.2025.20240153

摘要/Abstract

摘要：

在异质图数据收集中，由于隐私保护政策或版权限制，节点属性缺失现象普遍存在.针对属性不完备和属性完全缺失两种情况，提出了一种基于属性补全的异质图表示学习算法（HGAC）.对于属性不完备的节点，通过构建属性空间的邻接矩阵并执行图卷积来获取缺失的属性；将属性视为抽象节点，在元路径的引导下，对学习节点和属性进行拓扑嵌入，利用拓扑嵌入间的相似性来补全完全缺失的属性.在3个真实数据集上进行实验，结果表明，该算法有效提升了下游任务的性能，并具有较强的泛化能力.

关键词: 图表示学习, 异质图, 属性缺失, 属性补全, 元路径

Abstract:

In the process of collecting heterogeneous graph data， node attributes are often missing due to privacy protection policies or copyright constraints. Regarding both incomplete attributes and completely missing attributes， a heterogeneous graph representation learning algorithm based on attribute completion （HGAC） was proposed. For nodes with incomplete attributes， the missing attributes were obtained by constructing an adjacency matrix in the attribute space and performing graph convolution. Subsequently， the attributes were regarded as abstract nodes， and under the guidance of meta-paths， the topological embeddings of both nodes and attributes were learned. The similarity among the topological embeddings were then used to complete completely missing attributes. Experiments conducted on three real datasets demonstrate that the proposed algorithm effectively enhances the performance of downstream tasks and possesses strong generalization capability.

Key words: graph representation learning, heterogeneous graph, attribute missing, attribute completion, meta-path

中图分类号:

TP 391

陈东明, 刘嘉明, 梁春美, 王冬琦. 基于属性补全的异质图表示学习算法[J]. 东北大学学报（自然科学版）, 2025, 46(9): 25-33.

Dong-ming CHEN, Jia-ming LIU, Chun-mei LIANG, Dong-qi WANG. Heterogeneous Graph Representation Learning Algorithm Based on Attribute Completion[J]. Journal of Northeastern University(Natural Science), 2025, 46(9): 25-33.

图/表 4

参考文献 15

[1]	Chen F， Wang Y C， Wang B， et al. Graph representation learning： a survey［J］. APSIPA Transactions on Signal and Information Processing， 2020， 9： e15.
[2]	周丽华，王家龙，王丽珍，等. 异质信息网络表征学习综述［J］. 计算机学报， 2022， 45（1）： 160-189.
	Zhou Li-hua， Wang Jia-long， Wang Li-zhen， et al. Heterogeneous information network representation learning： a survey［J］. Chinese Journal of Computers， 2022， 45（1）： 160-189.
[3]	Jin D， Huo C Y， Liang C D， et al. Heterogeneous graph neural network via attribute completion［C］//Proceedings of the Web Conference 2021. Ljubljana， 2021： 391-400.
[4]	Kipf T N， Welling M. Semi-supervised classification with graph convolutional networks［C］//International Conference on Learning Representations. Toulon， 2017： 1-14.
[5]	Wang X， Ji H Y， Shi C， et al. Heterogeneous graph attention network［C］//The World Wide Web Conference. San Francisco， 2019： 2022-2032.
[6]	Fu X， Zhang J， Meng Z， et al. MAGNN： metapath aggregated graph neural network for heterogeneous graph embedding［C］//Proceedings of the Web Conference 2020. Taipei， 2020： 2331-2341.
[7]	You J， Ma X， Ding Y， et al. Handling missing data with graph representation learning［J］. Advances in Neural Information Processing Systems， 2020， 33： 19075-19087.
[8]	Taguchi H， Liu X， Murata T. Graph convolutional networks for graphs containing missing features［J］. Future Generation Computer Systems， 2021， 117： 155-168.
[9]	Chen X， Chen S， Yao J， et al. Learning on attribute-missing graphs［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2020， 44（2）： 740-757.
[10]	Tu W X， Zhou S H， Liu X W， et al. Initializing then refining： a simple graph attribute imputation network［C］//Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence. Vienna， 2022： 3494-3500.
[11]	Wang K， Yu Y， Huang C， et al. Heterogeneous graph neural network for attribute completion［J］. Knowledge-Based Systems， 2022， 251： 109171.
[12]	He D X， Liang C D， Huo C Y， et al. Analyzing heterogeneous networks with missing attributes by unsupervised contrastive learning［J］. IEEE Transactions on Neural Networks and Learning Systems， 2024， 35（4）： 4438-4450.
[13]	Zhu G， Zhu Z， Wang W， et al. Autoac： towards automated attribute completion for heterogeneous graph neural network［C］//2023 IEEE 39th International Conference on Data Engineering （ICDE）. Anaheim， 2023： 2808-2821.
[14]	Li C， Yan Y Y， Fu J H， et al. HetReGAT-FC： heterogeneous residual graph attention network via feature completion［J］. Information Sciences， 2023， 632： 424-438.
[15]	Dong Y X， Chawla N V， Swami A. Metapath2vec： scalable representation learning for heterogeneous networks［C］//Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Halifax， 2017： 135-144.

数据集	评价指标	训练比例	Metapath2vec	GCN	HAN	MAGNN	AC-HEN	HGNN-AC	HGAC
ACM	Macro-F1	20	69.43	72.31	90.01	88.01	90.09	90.87	91.49
		40	70.21	73.83	90.82	89.42	90.55	90.97	91.84
		60	70.57	72.59	91.51	90.39	91.22	91.13	92. 21
		80	71.74	72.23	91.71	90.79	91.39	91.52	92.77
	Micro-F1	20	71.13	75.27	89.89	88.08	90.43	90.75	91.63
		40	72.10	76.21	90.73	89.48	90.79	91.02	92.16
		60	72.85	76.10	91.37	90.42	91.23	91.61	92.45
		80	73.24	75.91	91.56	90.80	91.41	92.15	93.05
IMDB	Macro-F1	20	46.42	44.75	57.21	58.11	58.67	58.63	59.53
		40	47.70	45.26	57.54	59.39	59.11	59.39	59.77
		60	48.25	47.70	57.93	59.97	60.17	60.02	60.28
		80	48.73	48.25	58.16	60.02	60.34	60.57	60.84
	Micro-F1	20	48.08	47.44	57.24	58.16	58.66	58.72	59.69
		40	49.55	47.62	57.48	59.46	58.99	59.42	60.11
		60	50.06	48.49	57.68	60.05	59.87	59.84	60.71
		80	50.68	48.73	57.97	60.15	60.03	60.32	61.12
DBLP	Macro-F1	20	90.12	90.06	91.44	92.23	92.29	92.13	93.18
		40	90.74	90.37	91.62	92.87	92.79	92.74	93.73
		60	91.32	90.57	92.01	93.02	93.15	93.27	94.05
		80	91.68	90.74	92.15	93.18	93.22	93.37	94.47
	Micro-F1	20	91.25	90.53	91.89	92.35	92.43	92.58	92.94
		40	91.79	90.83	92.32	93.05	92.96	93.11	93.56
		60	92.33	91.01	92.59	93.38	93.28	93.52	93.98
		80	92.70	91.15	92.68	93.47	93.39	93.58	94.39

数据集	评价指标	训练比例	Metapath2vec	GCN	HAN	MAGNN	AC-HEN	HGNN-AC	HGAC
ACM	Macro-F1	20	69.43	72.31	90.01	88.01	90.09	90.87	91.49
		40	70.21	73.83	90.82	89.42	90.55	90.97	91.84
		60	70.57	72.59	91.51	90.39	91.22	91.13	92. 21
		80	71.74	72.23	91.71	90.79	91.39	91.52	92.77
	Micro-F1	20	71.13	75.27	89.89	88.08	90.43	90.75	91.63
		40	72.10	76.21	90.73	89.48	90.79	91.02	92.16
		60	72.85	76.10	91.37	90.42	91.23	91.61	92.45
		80	73.24	75.91	91.56	90.80	91.41	92.15	93.05
IMDB	Macro-F1	20	46.42	44.75	57.21	58.11	58.67	58.63	59.53
		40	47.70	45.26	57.54	59.39	59.11	59.39	59.77
		60	48.25	47.70	57.93	59.97	60.17	60.02	60.28
		80	48.73	48.25	58.16	60.02	60.34	60.57	60.84
	Micro-F1	20	48.08	47.44	57.24	58.16	58.66	58.72	59.69
		40	49.55	47.62	57.48	59.46	58.99	59.42	60.11
		60	50.06	48.49	57.68	60.05	59.87	59.84	60.71
		80	50.68	48.73	57.97	60.15	60.03	60.32	61.12
DBLP	Macro-F1	20	90.12	90.06	91.44	92.23	92.29	92.13	93.18
		40	90.74	90.37	91.62	92.87	92.79	92.74	93.73
		60	91.32	90.57	92.01	93.02	93.15	93.27	94.05
		80	91.68	90.74	92.15	93.18	93.22	93.37	94.47
	Micro-F1	20	91.25	90.53	91.89	92.35	92.43	92.58	92.94
		40	91.79	90.83	92.32	93.05	92.96	93.11	93.56
		60	92.33	91.01	92.59	93.38	93.28	93.52	93.98
		80	92.70	91.15	92.68	93.47	93.39	93.58	94.39

数据集	评价指标	Metapath2vec	GCN	HAN	MAGNN	AC-HEN	HGNN-AC	HGAC
ACM	NMI	21.22	51.40	61.37	63.17	64.82	64.54	65.22
ACM	ARI	21.00	53.01	64.39	67.41	68.84	68.26	69.75
IMDB	NMI	0.89	7.46	10.62	10.39	11.94	13.02	13.29
IMDB	ARI	0.22	7.69	10.01	11.11	12.17	13.43	13.87
DBLP	NMI	74.23	73.45	77.49	79.64	79.51	79.47	80.69
DBLP	ARI	78.11	77.50	82.95	82.80	84.51	84.66	84.95

数据集	评价指标	Metapath2vec	GCN	HAN	MAGNN	AC-HEN	HGNN-AC	HGAC
ACM	NMI	21.22	51.40	61.37	63.17	64.82	64.54	65.22
ACM	ARI	21.00	53.01	64.39	67.41	68.84	68.26	69.75
IMDB	NMI	0.89	7.46	10.62	10.39	11.94	13.02	13.29
IMDB	ARI	0.22	7.69	10.01	11.11	12.17	13.43	13.87
DBLP	NMI	74.23	73.45	77.49	79.64	79.51	79.47	80.69
DBLP	ARI	78.11	77.50	82.95	82.80	84.51	84.66	84.95