一种双通道半监督网络表示学习模型

doi:10.11959/j.issn.2096-0271.2024052

大数据 ›› 2024, Vol. 10 ›› Issue (4): 106-120.doi: 10.11959/j.issn.2096-0271.2024052

• 研究 • 上一篇

一种双通道半监督网络表示学习模型

杜航原¹, 谢富中¹, 王文剑¹, 白亮²

¹ 山西大学计算机与信息技术学院，山西太原 030006
² 山西大学智能信息处理研究所，山西太原 030006

出版日期:2024-07-01 发布日期:2024-07-01
作者简介:杜航原（1985- ），男，博士，山西大学计算机与信息技术学院副教授，主要研究方向为图机器学习理论及应用。
谢富中（1998- ），男，山西大学计算机与信息技术学院硕士生，主要研究方向为机器学习、网络数据挖掘。
王文剑（1968- ），女，博士，山西大学计算机与信息技术学院教授、院长，主要研究方向为机器学习、数据挖掘、计算智能。
白亮（1982- ），男，博士，山西大学智能信息处理研究所教授、所长，主要研究方向为机器学习与数据挖掘。
基金资助:
国家自然科学基金项目(U21A20513);国家自然科学基金项目(62076154);国家自然科学基金项目(62276159);国家自然科学基金项目(62276161);山西省重点研发计划项目(202302010101007);山西省重点研发计划项目(202202020101003);山西省基础研究计划项目(202303021221055)

A dual channel semi-supervised network representation learning model

Hangyuan DU¹, Fuzhong XIE¹, Wenjian WANG¹, Liang BAI²

¹ School of Computer and Information Technology, Shanxi University, Taiyuan 030006, China
² Institute of Information Processing, Shanxi University, Taiyuan 030006, China

Online:2024-07-01 Published:2024-07-01
Supported by:
The National Natural Science Foundation of China(U21A20513);The National Natural Science Foundation of China(62076154);The National Natural Science Foundation of China(62276159);The National Natural Science Foundation of China(62276161);The Key R＆D Program of Shanxi Province(202302010101007);The Key R＆D Program of Shanxi Province(202202020101003);The Fundamental Research Program of Shanxi Province(202303021221055)

摘要/Abstract

摘要：

在半监督网络表示学习中，节点标签对于网络在不同空间中映射关系的建立具有重要指导意义。然而在很多实际任务中，可用标签信息往往比较有限或难以获取，这导致在学习网络低维表示的过程中无法提供充分有效的监督。针对这一问题，提出了一种双通道半监督网络表示学习模型，该模型以自编码器为基本框架，由自监督和半监督两个信息传递通道构成。自监督信号与标签信息分别在两个通道中对网络表示映射关系的建立提供指导，同时二者之间形成信息互补与增强。考虑到两个通道间可能存在信息冗余，在互信息视角下设计了冗余识别与消除机制。在此基础上，构造了一体化优化模型，实现自监督学习与半监督学习的协同，使学习到的网络表示更好地捕捉和保持网络的结构和特性。在真实数据集上的实验结果表明，提出的模型学习的网络表示在节点分类、聚类和可视化等任务中能够获得优于基线方法的性能。

关键词: 半监督网络表示学习, 标签信息, 自监督学习, 互信息, 图神经网络

Abstract:

In semi-supervised network representation learning, node labels play an important role in guiding the establishment of network mapping relationships among different spaces.However, in many practical tasks, the available label information is usually limited or difficult to obtain, which makes it difficult to provide sufficient and effective supervision to the process of learning low-dimensional network representations.In order to solve this problem, a dual channel semisupervised network representation learning model is proposed, which is composed of two information transmission channels, namely self-supervised and semi-supervised channels, with the framework of autoencoder.The self-supervised information and the label information provide guidance for the establishment of network representation mapping in the two channels respectively, and they form such a sense of information complementation and enhancement for the learning process.Considering the possible information redundancy between the two channels, a redundancy recognition and elimination mechanism is designed in the perspective of mutual information.On this basis, an integrated optimization model is constructed to combine the self-supervised learning and the semi-supervised learning into a collaborative mechanism, which enables the learned representations to capture and preserve the structure and characteristics of the network.Experimental results on several real datasets show that the network representations learned by the proposed model can achieve better performance than baseline methods in node classification, clustering and visualization tasks.

Key words: semi-supervised network representation learning, label information, self-supervised learning, mutual information, graph neural network

中图分类号:

TP391

杜航原, 谢富中, 王文剑, 白亮. 一种双通道半监督网络表示学习模型[J]. 大数据, 2024, 10(4): 106-120.

Hangyuan DU, Fuzhong XIE, Wenjian WANG, Liang BAI. A dual channel semi-supervised network representation learning model[J]. Big Data Research, 2024, 10(4): 106-120.

图/表 10

图1

图2

表1

表2

不同数据集节点分类的实验结果"

方法	算法					数据集
		Cora		Citeseer		Pubmed		Facebook		Blogcatalog
		Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1
无监督	DeepWalk	70.09%± 0.85%	69.99%± 0.93%	69.55%± 0.54%	48.66%± 1.94%	74.67%± 1.16%	73.01%± 1.17%	64.32%± 1.67%	62.99%± 1.79%	33.47%± 0.61%	26.84%± 0.56%
	Node2vec	71.96%± 0.79%	71.86%± 0.80%	70.35%± 0.44%	50.37%± 1.12%	74.51%± 0.66%	73.10%± 0.63%	66.71%± 1.34%	65.16%± 1.53%	33.42%± 0.94%	26.37%± 0.86%
	DGI	81.47%± 0.20%	79.96%± 0.17%	72.11%± 0.38%	67.33%± 0.23%	77.13%± 0.18%	77.36%± 0.15%	71.29%± 0.14%	69.88%± 0.17%	44.12%± 0.63%	25.53%± 0.45%
	GMI	81.57%± 0.36%	80.42%± 0.45%	$73 . 12 % \pm 0 . 59 %$	66.37%± 0.56%	78.40%± 0.30%	78.32%± 0.35%	72.36%± 0.26%	70.55%± 0.34%	35.26%± 1.67%	23.14%± 1.29%
	GCN	81.02%± 0.09%	79.87%± 0.07%	69.55%± 0.54%	67.23%± 0.42%	78.38%± 0.26%	78.03%± 0.23%	68.17%± 0.05%	66.17%± 0.07%	45.92%± 1.45%	37.87%± 0.68%
有监督	GWNN	81.62%± 0.86%	80.25%± 0.79%	70.35%± 0.44%	67.32%± 0.29%	78.64%± 0.23%	77.98%± 0.26%	72.90%± 0.42%	$72 . 61 % \pm 0 . 41 %$	48.21%± 1.37%	38.45%± 1.17%
	GraphMLP	79.90%± 0.43%	78.04%± 0.49%	72.11%± 0.38%	68.16%± 0.36%	79.20%± 0.45%	78.66%± 0.48%	61.05%± 0.91%	59.93%± 0.80%	46.86%± 0.77%	36.77%± 0.53%
	DCSNRL	$83 . 57 % \pm 0 . 34 %$	$82 . 27 % \pm 0 . 37 %$	$73 . 12 % \pm 0 . 59 %$	$68 . 73 % \pm 0 . 59 %$	$80 . 85 % \pm 0 . 18 %$	$80 . 23 % \pm 0 . 16 %$	$73 . 41 % \pm 0 . 70 %$	72.22%± 0.73%	$53 . 37 % \pm 1 . 24 %$	$40 . 77 % \pm 0 . 36 %$

表2

表3

不同节点标记数量下各数据集节点分类的实验结果"

数据集				方法节点标记数量
		GCN		GWNN		GraphMLP		DCSNRL
		Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1	Micro-F1	Macro-F1
Cora	5	70.57%±1.03%	68.29%±0.98%	74.31%±0.64%	72.78%±1.05%	71.42%±1.78%	70.52%±1.70%	$76 . 38 % \pm 0 . 83 %$	$74 . 69 % \pm 0 . 92 %$
	10	73.91%±0.17%	72.96%±0.25%	75.50%±0.87%	74.70%±0.73%	74.48%±1.30%	72.71%±1.35%	$77 . 61 % \pm 0 . 46 %$	$76 . 45 % \pm 0 . 54 %$
	20	81.02%±0.09%	79.87%±0.07%	81.62%±0.86%	80.25%±0.79%	79.90%±0.43%	78.04%±0.49%	$83 . 57 % \pm 0 . 34 %$	$82 . 27 % \pm 0 . 37 %$
	30	82.76%±0.16%	81.74%±0.16%	82.96%±0.56%	81.82%±0.65%	81.64%±0.94%	80.07%±0.98%	$84 . 63 % \pm 0 . 23 %$	$83 . 52 % \pm 0 . 24 %$
	40	83.49%±0.24%	82.51%±0.28%	83.40%±0.57%	82.44%±0.71%	81.94%±0.33%	81.93%±0.33%	$84 . 98 % \pm 0 . 36 %$	$83 . 84 % \pm 0 . 37 %$
Citeseer	5	53.17%±0.22%	48.52%±0.32%	54.77%±1.73%	48.62%±2.57%	61.19%±1.67%	57.69%±1.94%	$65 . 88 % \pm 1 . 69 %$	$61 . 16 % \pm 1 . 39 %$
	10	66.51%±0.34%	63.68%±0.27%	65.92%±1.07%	62.62%±0.99%	65.70%±0.91%	62.25%±0.81%	$68 . 73 % \pm 0 . 54 %$	$64 . 91 % \pm 0 . 46 %$
	20	69.55%±0.54%	67.23%±0.42%	70.35%±0.44%	67.32%±0.29%	72.11%±0.38%	68.16%±0.36%	$73 . 12 % \pm 0 . 59 %$	$68 . 73 % \pm 0 . 59 %$
	30	72.17%±0.34%	69.37%±0.26%	72.56%±0.63%	69.73%±0.48%	71.14%±0.40%	67.56%±0.41%	$73 . 87 % \pm 0 . 48 %$	$69 . 85 % \pm 0 . 59 %$
	40	72.36%±0.27%	69.57%±0.32%	72.78%±0.28%	69.81%±0.78%	72.32%±0.65%	68.42%±0.61%	$73 . 91 % \pm 0 . 41 %$	$70 . 04 % \pm 0 . 45 %$
Pubmed	5	67.84%±0.19%	68.70%±0.27%	71.84%±0.05%	71.96%±0.05%	70.14%±2.24%	70.02%±2.11%	$75 . 01 % \pm 0 . 82 %$	$75 . 45 % \pm 0 . 73 %$
	10	71.43%±1.49%	71.97%±1.61%	72.76%±0.16%	72.72%±0.15%	72.06%±1.65%	72.43%±1.81%	$76 . 24 % \pm 0 . 67 %$	$76 . 34 % \pm 0 . 65 %$
	20	78.38%±0.26%	78.03%±0.23%	78.64%±0.23%	77.98%±0.26%	79.20%±0.45%	78.66%±0.48%	$80 . 85 % \pm 0 . 18 %$	$80 . 23 % \pm 0 . 16 %$
	30	80.40%±0.23%	79.65%±0.24%	79.14%±0.13%	78.44%±0.12%	79.34%±0.49%	78.89%±0.51%	$82 . 35 % \pm 0 . 61 %$	$81 . 53 % \pm 0 . 45 %$
	40	80.85%±0.52%	79.86%±0.51%	79.45%±0.12%	78.79%±0.08%	80.05%±0.48%	79.34%±0.57%	$82 . 55 % \pm 0 . 34 %$	$81 . 76 % \pm 0 . 27 %$
Facebook	5	57.80%±0.08%	55.81%±0.08%	62.30%±0.35%	60.79%±0.35%	47.04%±3.04%	46.12%±2.80%	$63 . 27 % \pm 1 . 20 %$	$61 . 55 % \pm 1 . 03 %$
	10	63.71%±0.24%	60.51%±0.22%	$69 . 21 % \pm 0 . 31 %$	$67 . 48 % \pm 0 . 24 %$	55.50%±1.93%	54.25%±1.87%	68.97%±1.16%	67.10%±1.08%
	20	68.17%±0.05%	66.17%±0.07%	72.90%±0.42%	$72 . 61 % \pm 0 . 41 %$	61.05%±0.91%	59.93%±0.80%	$73 . 41 % \pm 0 . 70 %$	72.22%±0.73%
	30	70.43%±0.47%	69.13%±0.07%	$74 . 62 % \pm 0 . 27 %$	73.34%±0.27%	62.48%±1.99%	61.22%±1.89%	74.37%±0.73%	$73 . 53 % \pm 0 . 65 %$
	40	75.14%±0.58%	74.15%±0.54%	78.20%±0.49%	77.43%±0.43%	66.71%±1.71%	65.43%±1.69%	$78 . 26 % \pm 0 . 67 %$	$77 . 54 % \pm 0 . 60 %$
Blogcatalog	5	38.83%±1.63%	30.08%±1.24%	40.12%±2.19%	32.21%±1.53%	$42 . 00 % \pm 1 . 57 %$	$33 . 02 % \pm 1 . 45 %$	39.73%±0.98%	30.24%±1.09%
	10	44.01%±1.69%	$36 . 62 % \pm 0 . 98 %$	45.60%±2.27%	35.18%±0.69%	43.94%±1.01%	34.61%±0.52%	$47 . 80 % \pm 0 . 17 %$	34.62%±0.46%
	20	45.92%±1.45%	37.87%±0.68%	48.21%±1.37%	38.45%±1.17%	46.86%±0.77%	36.77%±0.53%	$53 . 37 % \pm 1 . 24 %$	$40 . 77 % \pm 0 . 36 %$
	30	48.64%±1.36%	39.33%±0.91%	51.21%±2.24%	39.76%±1.21%	48.04%±0.87%	37.51%±0.61%	$58 . 32 % \pm 1 . 45 %$	$41 . 37 % \pm 1 . 65 %$
	40	48.83%±1.32%	40.44%±0.59%	53.10%±2.28%	40.98%±0.67%	48.81%±0.30%	38.65%±0.36%	$59 . 81 % \pm 1 . 10 %$	$41 . 53 % \pm 0 . 64 %$

表3

表4

不同数据集聚类的实验结果"

方法	算法					数据集
		Cora		Citeseer		Pubmed		Facebook		Blogcatalog
		NMI	ARI	NMI	ARI	NMI	ARI	NMI	ARI	NMI	ARI
无监督	DeepWalk	44.05%± 2.43%	33.76%± 4.01%	24.15%± 1.35%	11.22%± 0.45%	29.78%± 0.23%	29.83%± 0.36%	25.33%± 1.59%	17.92%± 1.94%	24.57%± 0.63%	16.67%± 0.93%
	Node2vec	44.55%± 2.64%	34.32%± 4.12%	24.91%± 0.41%	10.97%± 0.24%	29.84%± 0.23%	29.87%± 0.28%	25.24%± 1.31%	17.94%± 1.56%	24.39%± 0.53%	16.16%± 1.19%
	DGI	58.99%± 0.51%	55.54%± 1.19%	42.85%± 0.19%	40.71%± 0.29%	38.58%± 0.21%	36.77%± 0.22%	38.59%± 0.29%	34.19%± 0.60%	33.13%± 0.15%	23.01%± 0.19%
	GMI	58.26%± 1.12%	55.24%± 1.54%	41.44%± 0.58%	40.41%± 0.83%	33.03%± 2.66%	36.95%± 2.74%	40.67%± 1.51%	36.20%± 1.67%	28.17%± 1.56%	17.41%± 1.44%
有监督	GCN	41.04%± 0.68%	23.91%± 0.78%	37.97%± 0.36%	29.53%± 0.41%	34.16%± 0.22%	32.57%± 0.52%	36.77%± 0.46%	34.93%± 0.54%	28.94%± 0.80%	15.51%± 0.69%
	GWNN	44.19%± 1.09%	22.85%± 1.19%	36.43%± 1.29%	33.89%± 2.65%	37.38%± 0.18%	38.57%± 0.25%	$42 . 92 % \pm 0 . 57 %$	$41 . 66 % \pm 0 . 25 %$	28.57%± 0.57%	19.03%± 0.53%
	GraphMLP	50.12%± 0.51%	37.12%± 1.02%	41.96%± 0.83%	38.03%± 1.71%	36.04%± 0.69%	36.91%± 1.36%	25.54%± 2.55%	19.82%± 3.15%	39.96%± 0.81%	31.08%± 0.38%
	DCSNRL	$62 . 65 % \pm 0 . 26 %$	$62 . 99 % \pm 0 . 69 %$	$43 . 41 % \pm 0 . 26 %$	$43 . 78 % \pm 0 . 28 %$	$41 . 36 % \pm 0 . 57 %$	$46 . 32 % \pm 0 . 46 %$	42.17%± 0.60%	41.23%± 1.57%	$42 . 85 % \pm 0 . 16 %$	$36 . 51 % \pm 0 . 79 %$

表4

表5

不同节点标记数量下各数据集聚类的实验结果"

数据集	节点标记数量			方法
		GCN		GWNN		GraphMLP		DCSNRL
		NMI	ARI	NMI	ARI	NMI	ARI	NMI	ARI
Cora	5	29.11%±0.11%	14.23%±0.18%	32.17%±0.72%	13.49%±1.04%	46.71%±1.06%	34.64%±1.04%	$57 . 50 % \pm 0 . 47 %$	$57 . 00 % \pm 0 . 90 %$
	10	34.77%±0.10%	20.88%±0.28%	38.90%±1.26%	19.29%±0.69%	47.80%±1.26%	35.12%±0.87%	$58 . 04 % \pm 0 . 66 %$	$58 . 09 % \pm 0 . 71 %$
	20	41.04%±0.68%	23.91%±0.78%	44.19%±1.09%	22.85%±1.19%	50.12%±0.51%	37.12%±1.02%	$62 . 65 % \pm 0 . 26 %$	$62 . 99 % \pm 0 . 69 %$
	30	43.82%±1.84%	29.53%±1.98%	48.11%±1.01%	27.81%±0.28%	49.55%±0.44%	35.45%±1.02%	$63 . 57 % \pm 0 . 35 %$	$64 . 48 % \pm 0 . 56 %$
	40	45.31%±0.32%	27.54%±0.66%	60.08%±1.13%	42.81%±1.47%	50.18%±2.09%	39.32%±3.32%	$65 . 26 % \pm 0 . 44 %$	$66 . 15 % \pm 0 . 98 %$
Citeseer	5	27.85%±0.13%	20.87%±0.27%	25.19%±0.73%	17.49%±1.06%	36.65%±0.70%	32.32%±1.22%	$38 . 77 % \pm 0 . 92 %$	$38 . 50 % \pm 0 . 83 %$
	10	32.49%±0.20%	23.29%±0.93%	33.21%±2.14%	30.10%±2.46%	39.54%±2.12%	35.71%±3.26%	$41 . 98 % \pm 0 . 48 %$	$41 . 97 % \pm 0 . 78 %$
	20	37.97%±0.36%	29.53%±0.41%	36.43%±1.29%	33.89%±2.65%	41.96%±0.83%	38.03%±1.71%	$43 . 41 % \pm 0 . 26 %$	$43 . 78 % \pm 0 . 28 %$
	30	40.08%±0.07%	30.95%±0.16%	36.02%±1.05%	34.07%±1.87%	40.96%±0.79%	37.44%±0.83%	$44 . 65 % \pm 0 . 99 %$	$45 . 62 % \pm 1 . 71 %$
	40	41.87%±0.16%	35.31%±0.34%	39.48%±0.78%	34.94%±2.46%	42.66%±0.57%	38.71%±0.84%	$44 . 76 % \pm 0 . 16 %$	$45 . 70 % \pm 0 . 36 %$
Pubmed	5	30.39%±1.17%	30.03%±1.34%	31.41%±0.07%	31.69%±0.04%	30.27%±4.53%	25.67%±4.77%	$34 . 94 % \pm 1 . 11 %$	$38 . 06 % \pm 1 . 34 %$
	10	32.97%±0.70%	34.48%±0.54%	34.38%±0.16%	34.19%±0.13%	33.12%±1.24%	31.09%±1.80%	$36 . 47 % \pm 0 . 94 %$	$38 . 87 % \pm 1 . 02 %$
	20	34.16%±0.22%	32.57%±0.52%	37.38%±0.18%	38.57%±0.25%	36.04%±0.69%	36.91%±1.36%	$41 . 36 % \pm 0 . 57 %$	$46 . 32 % \pm 0 . 46 %$
	30	39.07%±0.31%	39.54%±0.65%	38.35%±0.15%	38.96%±0.07%	38.24%±0.73%	40.79%±0.89%	$44 . 48 % \pm 0 . 67 %$	$49 . 53 % \pm 0 . 57 %$
	40	40.16%±0.58%	41.79%±1.22%	38.42%±0.05%	39.27%±0.11%	39.89%±0.59%	41.82%±1.86%	$44 . 87 % \pm 0 . 45 %$	$51 . 02 % \pm 0 . 89 %$
Facebook	5	19.33%±0.78%	10.54%±1.03%	14.00%±0.71%	21.70%±0.35%	14.17%±2.91%	9.11%±2.54%	$29 . 72 % \pm 0 . 85 %$	$23 . 09 % \pm 0 . 59 %$
	10	30.15%±0.69%	25.14%±0.78%	35.97%±0.63%	$35 . 52 % \pm 0 . 40 %$	18.96%±3.46%	13.06%±2.21%	$36 . 26 % \pm 1 . 77 %$	34.03%±0.87%
	20	36.77%±0.46%	34.93%±0.54%	$42 . 92 % \pm 0 . 57 %$	$41 . 66 % \pm 0 . 25 %$	25.54%±2.55%	19.82%±3.15%	42.17%±0.60%	41.23%±1.57%
	30	39.03%±1.03%	37.64%±0.79%	$45 . 94 % \pm 0 . 74 %$	45.06%±0.37%	30.06%±2.17%	26.63%±3.83%	45.07%±0.32%	$45 . 50 % \pm 0 . 66 %$
	40	47.49%±0.82%	45.58%±1.21%	$51 . 16 % \pm 0 . 91 %$	$51 . 79 % \pm 1 . 02 %$	33.60%±2.64%	30.65%±2.61%	50.15%±0.41%	50.28%±1.14%
Blogcatalog	5	27.55%±0.83%	14.00%±1.27%	27.31%±0.45%	15.01%±1.03%	$33 . 89 % \pm 2 . 53 %$	24.59%±3.10%	30.62%±2.37%	$23 . 13 % \pm 2 . 48 %$
	10	27.71%±1.27%	14.81%±1.54%	27.18%±0.26%	15.05%±0.54%	34.34%±2.56%	24.85%±2.53%	$38 . 72 % \pm 0 . 52 %$	$27 . 03 % \pm 0 . 94 %$
	20	28.94%±0.80%	13.07%±0.71%	28.57%±0.57%	17.81%±0.61%	39.96%±0.81%	30.72%±1.34%	$42 . 85 % \pm 0 . 16 %$	$33 . 69 % \pm 1 . 35 %$
	30	28.95%±0.58%	13.17%±1.36%	29.23%±0.34%	18.73%±0.64%	41.43%±1.53%	29.09%±1.46%	$44 . 22 % \pm 0 . 16 %$	$36 . 22 % \pm 0 . 72 %$
	40	29.96%±0.53%	15.51%±0.69%	30.15%±0.41%	19.03%±0.53%	42.04%±1.12%	31.08%±0.38%	$44 . 31 % \pm 0 . 27 %$	$36 . 51 % \pm 0 . 79 %$

表5

表6

不同节点标记数量下消融实验结果"

数据集	评价指标	模型	节点标记数量
数据集	评价指标	模型	5	10	20	30	40
Cora	Micro-F1	DCSNRL_1	75.03%±1.12%	76.12%±0.78%	81.93%±0.83%	82.87%±0.69%	83.21%±0.84%
		DCSNRL	$76 . 38 % \pm 0 . 83 %$	$77 . 61 % \pm 0 . 46 %$	$83 . 57 % \pm 0 . 34 %$	$84 . 63 % \pm 0 . 23 %$	$84 . 98 % \pm 0 . 36 %$
	Macro-F1	DCSNRL_1	73.03%±1.30%	74.82%±0.93%	80.72%±0.81%	82.02%±0.53%	82.55%±0.71%
		DCSNRL	$74 . 69 % \pm 0 . 92 %$	$76 . 45 % \pm 0 . 54 %$	$82 . 27 % \pm 0 . 37 %$	$83 . 52 % \pm 0 . 24 %$	$83 . 84 % \pm 0 . 37 %$
	NMI	DCSNRL_1	54.67%±0.85%	55.43%±0.76%	60.19%±0.69%	61.12%±0.77%	62.35%±0.91%
		DCSNRL	$57 . 50 % \pm 0 . 47 %$	$58 . 04 % \pm 0 . 66 %$	$62 . 65 % \pm 0 . 26 %$	$63 . 57 % \pm 0 . 35 %$	$65 . 26 % \pm 0 . 44 %$
	ARI	DCSNRL_1	52.59%±1.27%	53.48%±1.03%	59.76%±0.85%	61.23%±0.91%	62.98%±0.82%
		DCSNRL	$57 . 00 % \pm 0 . 90 %$	$58 . 09 % \pm 0 . 71 %$	$62 . 99 % \pm 0 . 69 %$	$64 . 48 % \pm 0 . 56 %$	$66 . 15 % \pm 0 . 98 %$
Citeseer	Micro-F1	DCSNRL_1	64.46%±1.71%	67.30%±1.14%	70.22%±0.89%	71.65%±0.77%	72.00%±0.86%
		DCSNRL	$65 . 88 % \pm 1 . 69 %$	$68 . 73 % \pm 0 . 54 %$	$73 . 12 % \pm 0 . 59 %$	$73 . 87 % \pm 0 . 48 %$	$73 . 91 % \pm 0 . 41 %$
	Macro-F1	DCSNRL_1	60.65%±1.41%	63.23%±0.91%	66.34%±0.79%	67.59%±0.87%	68.11%±0.92%
		DCSNRL	$61 . 16 % \pm 1 . 39 %$	$64 . 91 % \pm 0 . 46 %$	$68 . 73 % \pm 0 . 59 %$	$69 . 85 % \pm 0 . 59 %$	$70 . 04 % \pm 0 . 45 %$
	NMI	DCSNRL_1	36.08%±1.06%	39.37%±0.88%	41.93%±1.12%	42.41%±1.31%	43.15%±0.94%
		DCSNRL	$38 . 77 % \pm 0 . 92 %$	$41 . 98 % \pm 0 . 48 %$	$43 . 41 % \pm 0 . 26 %$	$44 . 65 % \pm 0 . 99 %$	$44 . 76 % \pm 0 . 16 %$
	ARI	DCSNRL_1	35.67%±1.45%	39.21%±1.23%	40.89%±1.02%	43.00%±1.59%	43.74%±0.87%
		DCSNRL	$38 . 50 % \pm 0 . 83 %$	$41 . 97 % \pm 0 . 78 %$	$43 . 78 % \pm 0 . 28 %$	$45 . 62 % \pm 1 . 71 %$	$45 . 70 % \pm 0 . 36 %$

表6

表7

图3

参考文献 28

[1]	涂存超, 杨成, 刘知远 ,等. 网络表示学习综述[J]. 中国科学:信息科学, 2017,47(8): 980-996.
	TU C C , YANG C , LIU Z Y ,et al. Network representation learning:an overview[J]. Scientia Sinica (Informationis), 2017,47(8): 980-996.
[2]	戴筠 . 基于双曲空间图嵌入的科研热点预测[J]. 大数据, 2022,8(6): 94-104.
	DAI J . Emerging scientific topic prediction based on Poincare graph embedding[J]. Big Data Research, 2022,8(6): 94-104.
[3]	CAO S S , LU W , XU Q K . GraRep:learning graph representations with global structural information[C]// Proceedings of the 24th ACM International on Conference on Information and Knowledge Management. New York:ACM, 2015: 891-900.
[4]	PEROZZI B , AL-RFOU R , SKIENA S . DeepWalk:online learning of social representations[C]// Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. New York:ACM, 2014: 701-710.
[5]	WANG D X , CUI P , ZHU W W . Structural deep network embedding[C]// Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York:ACM, 2016: 1225-1234.
[6]	TU C C , ZHANG W C , LIU Z Y ,et al. Max-margin deepwalk:discriminative learning of network representation[C]// Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence. New York:ACM, 2016: 3889-3895.
[7]	KIPF T N , WELLING M . Semi-supervised classification with graph convolutional networks[EB]. arXiv preprint,2016,arXiv:1609.02907.
[8]	YANG Z L , COHEN W W , SALAKHUTDINOV R . Revisiting semi-supervised learning with graph embeddings[C]// Proceedings of the 33rd International Conference on International Conference on Machine Learning. New York:ACM, 2016: 40-48.
[9]	FAN S H , WANG X , SHI C ,et al. Debiased graph neural networks with agnostic label selection bias[J]. IEEE Transactions on Neural Networks and Learning Systems, 2022, 1-12.
[10]	ZHANG H G , WANG Z S , LIU D R . A comprehensive review of stability analysis of continuous-time recurrent neural networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2014,25(7): 1229-1262.
[11]	CHEN L Y , LI S B , BAI Q ,et al. Review of image classification algorithms based on convolutional neural networks[J]. Remote Sensing, 2021,13(22): 4712.
[12]	ZHANG G J , LIU Y , JIN X N . A survey of autoencoder-based recommender systems[J]. Frontiers of Computer Science, 2020,14(2): 430-450.
[13]	RUIZ L , GAMA F , RIBEIRO A . Gated graph recurrent neural networks[J]. IEEE Transactions on Signal Processing, 2020,68: 6303-6318.
[14]	GAO H Y , WANG Z Y , JI S W . Largescale learnable graph convolutional networks[C]// Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery ＆ Data Mining. New York:ACM, 2018: 1416-1424.
[15]	WENG Z Q , ZHANG W Y , DOU W . Adversarial attention-based variational graph autoencoder[J]. IEEE Access, 2020,8: 152637-152645.
[16]	CHENG D W , WANG X Y , ZHANG Y ,et al. Graph neural network for fraud detection via spatial-temporal attention[J]. IEEE Transactions on Knowledge and Data Engineering, 2022,34(8): 3800-3813.
[17]	ZHOU J , CUI G Q , HU S D ,et al. Graph neural networks:a review of methods and applications[J]. AI Open, 2020,1: 57-81.
[18]	WU Z H , PAN S R , CHEN F W ,et al. A comprehensive survey on graph neural networks[J]. IEEE Transactions on Neural Networks and Learning Systems, 2021,32(1): 4-24.
[19]	WU L R , LIN H T , TAN C ,et al. Selfsupervised learning on graphs:contrastive,generative,or predictive[J]. IEEE Transactions on Knowledge and Data Engineering, 2023,35(4): 4216-4235.
[20]	PENG Z , HUANG W B , LUO M N ,et al. Graph representation learning via graphical mutual information maximization[C]// Proceedings of The Web Conference 2020. New York:ACM, 2020: 259-270.
[21]	VELI?KOVI? P , FEDUS W , HAMILTON W L ,et al. Deep graph infomax[C]// Proceedings of the 7th International Conference on Learning Representations. arXiv preprint, 2019,arXiv:1809.10341.
[22]	SUN F , HOFFMANN J , VERMA V ,et al. InfoGraph:unsupervised and semisupervised graph-level representation learning via mutual information maximization[EB]. arXiv preprint,2019,arXiv:1908.01000.
[23]	BELGHAZI M I , BARATIN A , RAJESWAR S ,et al. MINE:mutual information neural estimation[EB]. arXiv preprint,2018,arXiv:1801.04062.
[24]	HJELM R D , FEDOROV A , LAVOIEMARCHILDON S ,et al. Learning deep representations by mutual information estimation and maximization[EB]. arXiv preprint,2018,arXiv:1808.06670.
[25]	QIN Z , KIM D , GEDEON T ,et al. Rethinking softmax with cross-entropy:neural network classifier as mutual information estimator[EB]. arXiv preprint,2018,arXiv:1911.10688.
[26]	HU Y , YOU H X , WANG Z C ,et al. Graph-MLP:node classification without message passing in graph[EB]. arXiv preprint,2021,arXiv:2106.04051.
[27]	XU B B , SHEN H W , CAO Q ,et al. Graph wavelet neural network[EB]. arXiv preprint,2019,arXiv:1904.07785.
[28]	KANG Z , LIN Z P , ZHU X F ,et al. Structured graph learning for scalable subspace clustering:from single view to multiview[J]. IEEE Transactions on Cybernetics, 2022,52(9): 8976-8986.

数据集	节点数	边数	特征维度	类别
Cora	2 708	5 429	1 433	7
Citeseer	3 327	4 732	3 703	6
Pubmed	19 717	44 338	500	3
Facebook	22 470	171 002	4 714	4
Blogcatalog	5 196	171 743	8 189	6

一种双通道半监督网络表示学习模型

A dual channel semi-supervised network representation learning model

在线阅读

PDF下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 10

参考文献 28

相关文章 2

Metrics

推荐阅读 0

[1]	兰海波. 混合型数据的邻域条件互信息熵属性约简算法[J]. 大数据, 2022, 8(4): 133-144.
[2]	王文广. 知识图谱推理：现代的方法与应用[J]. 大数据, 2021, 7(3): 42-59.