基于相似图投影学习的多视图聚类  

Multi-view clustering based on similarity graph projection learning

作　　者：赵伟豪 林浩申 曹传杰 杨晓君 Zhao Weihao;Lin Haoshen;Cao Chuanjie;Yang Xiaojun(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China;Unit 96901 of the PLA,Beijing 100094,China)

机构地区：[1]广东工业大学信息工程学院,广州510006 [2]中国人民解放军96901部队,北京100094

出　　处：《计算机应用研究》2024年第1期102-107,115,共7页Application Research of Computers

基　　金：广东省面上自然科学基金资助项目(2021A1515011141);国防重点实验室开放基金资助项目;国家自然科学基金青年资助项目(61904041)。

摘　　要：随着数据来源方式的多样化发展,多视图聚类成为研究热点。大多数算法过于专注利用图结构寻求一致表示,却忽视了如何学习图结构本身;此外,一些方法通常基于固定视图进行算法优化。为了解决这些问题,提出了一种基于相似图投影学习的多视图聚类算法(multi-view clustering based on similarity graph projection learning, MCSGP),通过利用投影图有效地融合了全局结构信息和局部潜在信息到一个共识图中,而不仅是追求每个视图与共识图的一致性。通过在共识图矩阵的图拉普拉斯矩阵上施加秩约束,该算法能够自然地将数据点划分到所需数量的簇中。在两个人工数据集和七个真实数据集的实验中,MCSGP算法在人工数据集上的聚类效果表现出色,同时在涉及21个指标的真实数据集中,有17个指标达到了最优水平,从而充分证明了该算法的优越性能。With the diversified development of data sources,multi-view clustering has become a research hotspot.Most algorithms focus too much on using graph structure to seek consistent representation,but ignore how to learn the graph structure itself.In addition,some methods are usually optimized based on fixed views.In order to solve these problems,this paper proposed a multi-view clustering algorithm based on similarity graph projection learning(MCSGP),which effectively fused the global structure information and local potential information into a consensus graph by using the projection graph,rather than only pursuing the consistency of each view with the consensus graph.By imposing a rank constraint on the graph Laplacian matrix of the consensus graph matrix,this algorithm could naturally divide the data points into the required number of clusters.In the experiments on two artificial datasets and seven real datasets,the MCSGP algorithm shows excellent clustering effect on artificial data sets.At the same time,in the real datasets involving 21 indicators,17 indicators reach the optimal level,which fully proves the superior performance of the proposed algorithm.

关 键 词：多视图聚类 投影学习 相似图 图融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]