全局与局部结构学习的多视图子空间聚类算法

Global and Local Structure Learning for Multi-view Subspace Clustering

约束双线性分解的多视图子空间聚类算法(CBF-MSC)忽略了视图局部结构信息,导致信息损失,进而影响多视图聚类效果。针对上述问题,提出了全局与局部结构学习的多视图子空间聚类算法(CBF-LGLS)。该算法首先考虑了视图的一致性与互补性,认为不同视图的系数矩阵应该具有相同的聚类属性,而不是在多个视图之间是一致的,从而充分探索挖掘视图底层数据分布和聚类属性。该算法还全面考虑了视图的局部结构信息,有效捕获单个视图的内在差异,减少了信息损失。此外,该算法采用了自适应加权的方法,减少了噪声与冗余对聚类效果的影响。对于每个视图预定义相似度矩阵的传统模式,采用了自适应距离正则化方法,达到充分考虑单个视图的几何结构与视图之间相同的簇结构的目的,进而提高聚类效果。算法在广泛使用的数据集上进行实验,并与主流算法进行比较,结果表明,提出的算法具有良好的聚类效果和收敛性。

Constrained bilinear factorization multi-view subspace clustering algorithm(CBF-MSC)ignores local structure information of views,resulting in loss of information and thus affecting the effect of multi-view clustering.To solve the above problems,a multi-view subspace clustering algorithm based on global and local structure learning(constrained bilinear factorization by learning global and local structures for multi-view subspace clustering,CBFLGLS)is proposed.The algorithm considers the consistency and complementarity of views,and considers that the coefficient matrices of different views should have the same clustering attributes,rather than being consistent among multiple views,so as to fully explore the underlying data distribution and clustering attributes of mining views.Moreover,the algorithm also fully considers the local structure information of the view,effectively captures the internal differences of a single view and reduces the loss of information.Then,the algorithm comprehensively considers the local structure information of views,effectively capturing the internal differences of a single view and reducing the loss of information.In addition,the algorithm adopts an adaptive weighting method to reduce the impact of noise and redundancy on clustering effect.For the traditional pattern of predefined similarity matrices for each view,an adaptive distance regularization method is adopted to fully consider the geometric structure of a single view and the same cluster structure between views,thereby improving clustering performance.The algorithm is tested on widely-used datasets and compared with mainstream algorithms.The results show that the proposed algorithm obtains good clustering effect and convergence.