基于局部约束线性编码的多视角聚类方法

A Multi-view Clustering Method with Locality-constrained Linear Coding

由于多视角数据能够提供更多有用信息来提高聚类效果,多视角聚类得到了广泛的研究。基于子空间的多视角聚类方法是流行的研究方法,主要是从多视角数据中学习一个统一表示,用于聚类。现有的基于子空间的多视角学习方法具有不错的聚类效果,但往往忽略了样本之间的局部结构信息。因此,提出了一种新的多视角学习方法,该方法采用局部约束线性编码来获取样本之间的局部结构,提高所学到的表示的质量。同时,利用多视角数据的数据特性,每个视角的表示由视角共享部分和视角独有部分组成,更充分挖掘数据所存在的有效信息。另外,采用图正则约束来保持数据的内在流形结构,提高所学到表示的质量。最后,在真实数据集上的实验结果表明,所提出的方法优于现有的多视角聚类方法。

Multi-view clustering has been widely studied as multi-view data can provide more useful information for improving the clustering.Multi-view clustering methods based on subspace are popular methods,which mainly learn a unified representation from multi-view data for clustering.The existing multi-view learning methods based on subspace have good clustering effect,but ignore the local structure information between samples.Thus,this paper proposes a new multi-view learning method.The locality-constrained linear coding is used to obtain local structures between samples to improve the quality of the learned representations.At the same time,utilizing the property of multi-view data,the representation of each view is composed of shared part and view-specific part,so as to fully mine the effective information from data. In addition,graph regularization constraint is used to preserve the intrinsic manifold structure of data for improving the quality of the learned representations.Finally,experimental results on real dataset show that the proposed method is superior to existing multi-view clustering methods.