摘要
边分类是图挖掘和社交网络分析中的一个重要研究方向。然而,现有边分类方法往往通过聚合连边端点的表示来间接提取边的特征,且设计为非端到端的学习方式,会造成较大的信息损失。针对上述问题,该文提出一种融合全局和局部信息的边分类模型EGLec(end-to-endModel withglobal andlocal InformationFusionfor edge classification),将边特征提取和边分类过程建模成端到端的训练方式。首先,根据网络中所有节点对边的权重构建边的全局信息。其次,结合图自编码器和深度自编码器,分别提取网络的结构特征和边全局信息的深层语义特征,以生成连边的结构嵌入和全局特征嵌入。最后,融合结构嵌入和全局特征嵌入得到最终的连边表示以用于边分类。在三个真实数据集上的对比实验验证了该文所提出模型可显著提高边分类性能。
Edge classification is an important research topic in the field of graph mining and social network analysis.In contrast to the non-end-to-end existing methods aggregated from nodes,we propose EGLec,i.e.an End-to-end model with Global and Local information fusion for Edge Classification,to combine the edge feature extraction and the edge classification.Specifically,we quantify global information of edges according to contribution of all nodes in the network to edges.Combined with graph auto-encoder and deep auto-encoder,we extract structure information of network and capture high-level feature of global information to generate the structure and the global feature embed-dings for edges,respectively.After that,the structure embedding and global feature embedding are merged to obtain final edge representation for edge classification.Experiments on three real-world social networks demonstrate the effectiveness compared with baselines.
作者
颜登程
储蓓
张以文
李冬
倪莉
YAN Dengcheng;CHU Bei;ZHANG Yiwen;LI Dong;NI Li(School of Computer Science and Technology,Anhui University,Hefei,Anhui 230601,China;School of Computer Engineering,Anhui Wenda University of Information Engineering,Hefei,Anhui 231201,China)
出处
《中文信息学报》
CSCD
北大核心
2023年第4期137-145,共9页
Journal of Chinese Information Processing
基金
国家重点研发计划(2019YFB1704101)
国家自然科学基金(61872002)
安徽省自然科学基金(2208085QF197)
安徽省高校自然科学研究重点项目(2022AH05008637)。
关键词
边分类
网络表示学习
自编码器
edge classification
network representation learning
auto-encoder
