共享和特定表示的多视图属性图聚类  

Multi-view Attributed Graph Clustering Based on Shared and Specific Representation

作　　者：曹付元[1] 陈晓惠 CAO Fu-Yuan;CHEN Xiao-Hui(School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,山西太原030006

出　　处：《软件学报》2025年第3期1254-1267,共14页Journal of Software

基　　金：国家自然科学基金(61976128);山西省科技创新人才团队(202204051002016)。

摘　　要：现有的多视图属性图聚类方法通常是在融合多个视图的统一表示中学习一致信息与互补信息,然而先融合再学习的方法不仅会损失原始各个视图的特定信息,而且统一表示难以兼顾一致性与互补性.为了保留各个视图的原始信息,采用先学习再融合的方式,先分别学习每个视图的共享表示与特定表示再进行融合,更细粒度地学习多视图的一致信息和互补信息,构建一种基于共享和特定表示的多视图属性图聚类模型(multi-view attribute graph clustering based on shared and specific representation,MSAGC).具体来说,首先通过多视图编码器获得每个视图的初级表示,进而获得每个视图的共享信息和特定信息;然后对齐视图共享信息来学习多视图的一致信息,联合视图特定信息来利用多视图的互补信息,通过差异性约束来处理冗余信息;之后训练多视图解码器重构图的拓扑结构和属性特征矩阵;最后,附加自监督聚类模块使得图表示的学习和聚类任务趋向一致.MSAGC的有效性在真实的多视图属性图数据集上得到了很好地验证.Existing multi-view attributed graph clustering methods usually learn consistent information and complementary information in a unified representation of multiple views.However,not only will the specific information of the original views be lost under the method of learning after fusion,but also the consistency and complementarity are difficult to balance under the unified representation.To retain the original information of each view,this study adopts the method of learning first and then fusing.Firstly,the shared representation and specific representation of each view are learned separately before fusion,and the consistent information and complementary information of multiple views are learned more fine-grained.A multi-view attributed graph clustering model based on shared and specific representation(MSAGC)is constructed.Specifically,the primary representation of each view is obtained by a multi-view graph encoder,and then the shared information and specific information of each view are obtained.Then the consistent information of multiple views is learned by aligning the view shared information,the complementary information of multiple views is utilized by combining the view specific information,and the redundant information is processed through the difference constraint.After that,the topological structure and attribute feature matrix of the multi-view decoder reconstruction graph are trained.Finally,the additional self-supervised clustering module makes the learning and clustering tasks of graph representation tend to be consistent.The effectiveness of MSAGC is well verified on real multi-view attributed graph datasets.

关 键 词：多视图属性图 共享信息 特定信息 聚类 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]