基于自表示和投影映射的不完整多视图聚类

Incomplete Multi-view Clustering Based on Self-representation and Projection Mapping

针对不完整多视图聚类存在的缺陷,提出一种融合自表示和投影映射的统一框架.首先,利用自表示和样本存在指示矩阵学习一致相似图,它反映了样本间的公共相似关系;其次,利用投影映射将样本矩阵投影到超球面上,得到公共低维表示;最后,将两者通过谱表示嵌入在一起,解决了因多视图数据缺失引起的不完整多视图聚类问题.该算法在真实数据集上的实验结果优于其他算法,证明了算法的有效性.

Aiming at the shortcomings of incomplete multi-view clustering,we proposed a unified framework that integrated self-representation and projection mapping.Firstly,self-representation and sample presence indication matrices were used to learn a uniform similarity graph,which reflected the common similarity relationship between samples.Secondly,the sample matrices were projected onto the hypersphere by using projection mapping to obtain a common low-dimensional representation.Finally,the two were embedded together through spectral representation to solve the incomplete multi-view clustering problem caused by missing multi-view data.The experimental results of this algorithm on real datasets are better than other algorithms,which proves the effectiveness of the proposed algorithm.