块对角引导的多视角统一图聚类

Multi-view Unified Graph Clustering Guided by Block Diagonal

多视角聚类能够整合多个视角的信息来提高聚类效果,目前很多研究都限于关注多视角一致性,得到的统一相似度图中仍存在许多非同簇之间的关系,甚至当某些簇的噪声达到一定程度时还可能导致统一相似度图难以形成簇的块对角结构.为此,本文提出一种块对角引导的多视角统一图聚类方法,该方法先将不同视角的相似度图分解成一致性部分与不一致性部分;然后通过构造不一致性关系来获得更纯净的一致性部分;进而融合所有视角的一致性部分建立一个相似度图;最后在该相似度图中加入块对角引导和连通分量约束,学习到高质量的统一相似度图.通过在六个数据集上进行对比实验,证明了本文提出的方法的有效性.

Multi-view clustering can integrate information from multiple view to promote clustering performance.Unfortunately,many studies are limited to focusing on multi-view consistency and there exist many relationships between different clusters in unified graph.What is worse,when the noise of some clusters reaches a certain level,it may lead the unified graph to be difficult to form the cluster block diagonal structure.To address issues mentioned above,a model is proposed in this paper,it firstly decomposed the similarity graphs of each view into consistent part and inconsistent part respectively.In order to obtain purer consistent parts of each view,it con-structed multi-view inconsistency.Then a similarity graph is obtained by fusing the consistent parts of all views.At the same time,the block diagonal-guided and connected component constraints are imposed on the graph.Finally,we obtain a better unified graph.Exten-sive experiments on six datasets demonstrate that the effectiveness of the proposed method.