利用邻域k元节点组信息的节点结构相似性判定方法
A Structure Similarity Determination Method for Aggregation Label Information of k-tuple Group in Node Neighborhood
杨贵 1韦兴宇 1郑文萍2
作者信息
- 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 计算机与信息技术学院,山西 太原 030006;计算智能与中文信息处理教育部重点实验室(山西大学),山西 太原 030006;山西大学 智能信息处理研究所,山西 太原 030006
- 折叠
摘要
复杂网络中的节点往往形成一些频繁出现且具有特定局部连接模式的高阶子结构,利用这些高阶子结构可以更好地刻画网络的拓扑特征及相关功能模块.通过度量节点间的结构相似性,有助于研究拓扑结构中节点之间的交互模式,理解复杂网络的局部结构和功能.为更充分利用节点邻域的高阶结构信息,提出了一种利用节点邻域内的k元节点组标签信息的结构相似性判定方法GANNLI(Group-based Aggregated Neighborhood Label Infor-mation).该方法首先将k元节点组形成的非同构子图作为其组标签,再利用WL方法对k元节点组的邻域组标签信息进行聚合和更新,统计节点所构成的不同k元组的标签信息以得到节点表示,并利用余弦相似度计算节点间的结构相似性.与仅考虑节点度、接近中心性等低阶信息的方法相比,本方法利用高阶k元组结构信息更有效地度量了节点间的结构相似性.在真实网络数据集上的实验结果表明,所提出的GANNLI算法能更有效地计算节点间的结构相似性,在节点分类任务中的性能相比Struc2vec提高了2%至6%,相比Node2vec提高了8%至14%.
Abstract
In complex networks,nodes often form higher-order substructures with specific local connectivity patterns that frequently appear.These higher-order substructures can better characterize the network's topological features and related functional modules.Measuring the structural similarity between nodes aids in studying the interaction patterns within the network's topology and under-standing the local structures and functions of complex networks.To fully utilize the higher-order structural information in node neighborhoods,we propose a method for determining structural similarity using the label information of k-tuple groups in node neighborhoods,called GANNLI(Group-based Aggregated Neighborhood Label Information).This method first forms non-isomor-phic subgraphs as group labels for k-tuple node groups,and then uses the Weisfeiler-Lehman(WL)method to aggregate and update the neighborhood group label information of k-tuple node groups.It counts the label information of different k-tuples formed by nodes to obtain node representations and calculates the structural similarity between nodes using cosine similarity.Compared to methods that only consider low-order information such as node degree and closeness centrality,our approach leverages higher-order k-tuple structural information to more effectively measure structural similarity between nodes.Experimental results on real network datasets demonstrate that the proposed GANNLI algorithm can more effectively calculate structural similarity between nodes,there-by improving the performance of node classification tasks.Specifically,the GANNLI shows a performance improvement of 2%to 6%over Struc2vec algorithm and an improvement of 8%to 14%over Node2vec algorithm.These results indicate that the GANN-LI's ability to incorporate higher-order structural information into the analysis of node neighborhoods allows for more accurate and insightful modeling of complex networks,leading to enhanced understanding and better performance in practical applications.
关键词
复杂网络/结构相似性/k元组/高阶结构/Weisfeiler-Lehman方法Key words
complex network/structural similarity/k-tuple/higher-order structure/Weisfeiler-Lehman引用本文复制引用
基金项目
国家自然科学基金(62072292)
山西省1331工程项目()
出版年
2024
国家科技期刊平台
NSTL
万方数据