基于一步张量学习的多视图子空间聚类

One-step Tensor Learning for Multi-view Subspace Clustering

现有多视图子空间聚类算法通常先进行张量表示学习,进而将学习到的表示张量融合为统一的亲和度矩阵.然而,因其独立地学习表示张量和亲和度矩阵,忽略了两者之间的高度相关性.为了解决此问题,提出一种基于一步张量学习的多视图子空间聚类方法,联合学习表示张量和亲和度矩阵.具体地,该方法对表示张量施加低秩张量约束,以挖掘视图的高阶相关性.利用自适应最近邻法对亲和度矩阵进行灵活重建.使用交替方向乘子法对模型进行优化求解,通过对真实多视图数据的实验表明,较于最新的多视图聚类方法,提出的算法具有更好的聚类准确性.

A surge of the existing multi-view subspace clustering algorithms generally learn the third-order tensor representation first and then fuse the learned representation tensor into a unified affinity matrix.However,since they learn the representation tensor and the affinity matrix independently,they cannot seamlessly capture their high-order correlation.To address this challenge,we propose a novel multi-view subspace clustering method based on one-step tensor learning(OTSC)to jointly learn the representation tensor and affinity matrix.Specifically,we impose the low-rank tensor constraint on the representation tensor to explore the correlation of high-order crossviews dexterously,utilize the adaptive nearest neighbor strategy to reconstruct a flexible affinity matrix,and adopt the alternating direction method of multipliers(ADMM)to optimize our model.Extensive experiments on real multi-view data demonstrated the superiority of OTSC compared to the state-of-the-art methods.