基于图表示学习的社交网络群体竞争影响力识别
Group competitive influence identification of social network groups based on graph representation learning
刘鑫哲 1方勇 1贾鹏 1寇蒋恒 1范希明 1周小涵 1潘睿 1朱旭2
作者信息
- 1. 四川大学网络空间安全学院,成都 610207
- 2. 国家能源集团新疆能源有限责任公司,乌鲁木齐 830000
- 折叠
摘要
群体竞争影响力识别是社交网络分析领域的一个必要研究,其任务是识别社交网络中任意两群体节点在相互竞争条件下的影响力,在舆情分析等实际场景中具有重要意义.在过去的几年里,许多研究集中在没有竞争对手的群体影响力识别.然而,竞争普遍存在于真实的社交网络中,因此研究群体竞争影响力识别任务十分必要.与非竞争场景下的群体影响力识别相比,群体竞争影响力识别存在竞争数据集的构建和群体对嵌入聚合等挑战.图表示学习(GRL)在社交网络分析领域取得了巨大的成功,可以将图结构表示成具有结构信息的低维嵌入,能够更好的反应节点之间的相互作用,提供比传统方法更丰富的信息.本文开创性的使用GRL来解决竞争场景下的群体影响力识别问题,并提出了一个基于GRL的框架.为了解决竞争数据集的构建问题,本文提出了一种基于影响力多样性的群体对构建方法.为了解决竞争群体对嵌入聚合问题,本文提出了一种基于求和相减的方法来聚合竞争群体对中节点的嵌入.本文在7个真实的社交网络上进行了大量实验来分析所提框架的有效性.实验结果表明所提框架优于基线方法.
Abstract
Group competitive influence identification is a necessary area in social network analysis,whose task is to identify the influence of two groups of nodes in a social network under competing conditions,and has significant implications in practical scenarios such as public opinion analysis.In recent years,much re-search has focused on identifying group influence without competitors.However,competition is ubiquitous in real-world social networks,making it necessary to study group competitive influence identification tasks.Compared to group influence identification in non-competitive scenarios,group competitive influence identifi-cation presents challenges such as the construction of competitive datasets and embedding aggregation of com-petitive groups.Graph representation learning(GRL)has been successful in social network analysis,as it can represent graphical structure as low-dimensional embeddings with structural information and interactions be-tween nodes,providing richer information than traditional methods.This paper pioneered the use of GRL to solve group influence prediction problems in competitive scenarios and proposed a GRL-based framework.To address the problem of constructing competing datasets,a group pair construction method based on influence diversity is proposed.To address the problem of embedding aggregation,a method based on summation-subtraction is proposed to aggregate the embeddings of nodes in a competitive group pair.Extensive experi-ments are conducted on seven real social networks to analyze the effectiveness of the proposed framework.The experimental results show that the proposed framework outperforms the baseline approach.
关键词
群体竞争影响力识别/社交网络分析/深度学习/图神经网络Key words
Group competitive influence identification/Social network analysis/Deep learning/Graph neural network引用本文复制引用
基金项目
四川省科技厅重点研发项目(2021YFG0156)
出版年
2024
国家科技期刊平台
NSTL
万方数据