基于合并子图的双通道跨网络用户身份识别  

Two-channel cross-network user identity linkage based on merged subgraph

作　　者：周小涵 贾鹏 杨频 寇蒋恒 刘鑫哲 ZHOU Xiao-Han;JIA Peng;YANG Pin;KOU Jiang-Heng;LIU Xin-Zhe(School of Cyber Science and Engineering,Sichuan University,Chengdu 610207,China)

机构地区：[1]四川大学网络空间安全学院,成都610207

出　　处：《四川大学学报（自然科学版）》2024年第4期3-13,共11页Journal of Sichuan University(Natural Science Edition)

基　　金：四川省科技厅重点研发项目(2021YFG0156)。

摘　　要：跨社交网络的用户身份识别(UIL)的本质是通过各种方法发现跨社交平台上的同一用户或者实体.现有方法的最新思路主要是对网络中节点的各种结构或属性特征进行聚合,然后构建相应的深度学习模型,学习相同用户在不同网络中特征的相似性,以此来实现不同网络中相同用户的对齐.但是大多数方法较少考虑用户的属性信息或者是只用单一方法来处理不同类型的属性特征,这样处理的后果就是不能完美捕获到属性文本中的有效特征.此外,现有的方法是对2个网络分别在各自的嵌入空间进行学习然后映射到同一个公共空间,也就只能学习到各自网络的信息.本文提出了一个新的方法,即基于合并子图的双通道跨网络用户身份识别(TCUIL).为了解决获取节点特征单一性问题,提出了多维特征提取方法实现了针对不同特征采用不同方法进行处理.为了解决2个网络嵌入空间互不相交的问题,提出了图合并方法实现了2个网络中信息的交互.此外,为了能学习到2个网络的多维度信息,设计了双通道网络结构实现了对网络拓扑结构、属性特征、节点间关系特征的有效学习.通过在2个真实数据集上的大量实验,证明了本文方法优于现有最先进的对齐方法 .我们在2个真实数据集(社交网络和合著网络)上进行了大量实验,在F1方面社交网络至少提高了44.32%,合著网络至少提高了25.04%.The essence of cross-social network user identity linkage(UIL)is to discover the same user or entity across different social platforms through various methods.The latest idea of existing methods is mainly to aggregate various structural or attribute features of nodes in the network,and then build corresponding deep learning models to learn the similarity of features for the same user in different networks,thereby achieving alignment of the same user in different networks.However,most methods rarely consider user attribute information or only use a single method to handle different types of attribute features,resulting in the inability to perfectly capture the effective features in attribute texts.Additionally,existing methods learn from the embedding space of two networks separately and then map them to a common space,which only captures information from each network.This paper proposes a new method,namely Two-Channel Cross-Network User Identity Linking Based on Merged Subgraphs(TCUIL).To address the issue of obtaining singular node features,a multi-dimensional feature extraction method is proposed to handle different features using different methods.To solve the problem of non-intersecting embedding spaces in two networks,a graph merging method is proposed to facilitate interaction between the information in the two networks.Furthermore,to learn multi-dimensional information from the two networks,a two-channel network structure is designed to effectively learn the network topology,attribute features,and inter-node relationship features.Through extensive experiments on two real datasets,the proposed method has been shown to outperform state-of-the-art alignment methods.We conducted extensive experiments on two real datasets(social network and co-authored network),achieving at least a 44.32%improvement in F1 for the social network and at least a 25.04%improvement for the co-authored network.

关 键 词：社交网络 用户身份识别 合并子图 图神经网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]