非线性多视角子空间聚类方法

不同视角特征构成的数据比单视角特征具有更多的信息,充分利用多视角特征可以提高聚类效果.由于不同视角空间中的特征不具有可比性,基于线性表示理论的子空间学习方法通过学习表示矩阵挖掘互补信息.但现实数据多是非线性的,线性表示理论不利于发现数据的非线性关系.针对该问题,采用非线性投影及流形正则项来刻画多视角下的非线性数据,实验结果表明,所提方法能够对多视角数据进行有效聚类.

Nonlinear multi-view subspace clustering methods

Data composed of different view features have more information than single view features, and making full use of multiple view feature can improve the clustering performance. Since features from different viewing space are not comparable, the subspace learning method based on linear representation theory mines complementary information by learning representation matrix. However, most of the real data are nonlinear, and the linear representation theory is not conductive to finding the relationship of the data. To solve this problem, the nonlinear projection and manifold regularization terms are adopted to describe the nonlinear data from multiple views. Experimental results show that the proposed method can effectively cluster the multi-view data.