摘要
基于图的多视图聚类算法通过探索样本点之间的邻近关系,受到了广泛的关注.尽管在实际应用中已经取得了较好的聚类性能,但是观察到大多数算法只是利用一阶邻近关系去构建相似图,这导致信息探索不足和多视图数据表征能力下降.为了解决这个挑战,本文提出了一种新颖的基于高阶图融合的多视图聚类算法(high-order graph fusion for multi-view clustering,HCDMC).具体地,所提出的算法通过一种新颖的隐式权重学习策略,从每个视图对应的一阶和二阶邻近图中学习相应的高阶图.引入希尔伯特-施密特(Hilbert-Schmidt)独立性准则作为一种差异性正则化项,旨在加强一致性高阶图的互补信息.最后,对学习到的一致性高阶图施加连通性约束,直接得到聚类标签矩阵,无需任何后处理步骤.使用交替方向乘子法去解决模型的优化问题.在6个真实的数据集上进行了一系列的实验,相较于最新的算法,本文提出的算法具有更好的聚类性能.
Graph-based multi-view clustering has received considerable attention due to its ability to explore neighborhood relationships among data points from multiple views.Although these methods have achieved impressive clustering performance and efficiency in various applications,most of them only merely exploit the first-order proximity graph within multi-view data,which results in inadequate information exploration and a degraded capacity of multi-view representation.In this paper,we propose a high-order graph fusion for multiview clustering(HCDMC).Specifically,the proposed algorithm constructs the corresponding high-order graph from first-order and second-order proximity graphs for each view by a novel implicit weight learning paradigm.Then,Hilbert Schmidt Independence Criterion as a diversity regularization term is to improve the complementary information of the consensus high-order graphs.Finally,we impose a connectivity constraint on the consensus high-order graph to obtain clustering labels directly without any post-processing step.An efficient iterative algorithm with good convergence is designed to solve the resulting optimization problem.Extensive experiments conducted on six multi-view datasets demonstrate the promising performance of our proposed when compared to recent state-of-the-art algorithms.
作者
尤运宁
唐厂
刘新旺
邹鑫
刘袁缘
蒋良孝
张长青
Yunning YOU;Chang TANG;Xinwang LIU;Xin ZOU;Yuanyuan LIU;Liangxiao JIANG;Changqing ZHANG(School of Computer Science,China University of Geosciences,Wuhan 430074,China;School of Computer Science,National University of Defense Technology,Changsha 410073,China;College of Intelligence and Computing,Tianjin University,Tianjin 300350,China)
出处
《中国科学：信息科学》
CSCD
北大核心
2024年第9期2098-2115,共18页
Scientia Sinica(Informationis)
基金
国家自然科学基金面上项目(批准号:62076228,62476258)
国家杰出青年科学基金项目(批准号:62325604)资助。
关键词
多视图聚类
高阶图
图结构学习
图融合
差异性正则化
multi-view clustering
high-order graph
graph structure learning
graph fusion
diversity regularization
