Self-representation-based Subspace Representation For Semi-supervised Learning

Many real-world data are high-dimensional,which often lie in the sturcture of lowdimensional underlying subspaces.Self-representation based subspace learning has been successfully applied to clustering tasks for high-dimensional data,in which the key assumption is that data are from multiple subspaces and can be reconstructed by the data themselves.In reality,it is difficult to collect numerous data with labels.Compared to supervised learning,semi-supervised learning can get rid of the dependence on labeled samples.Simultaneously,incorporating a small amount of supervised information can guide unsupervised learning to construct a more accurate model.For the problem of label shortage on high-dimensional data,semi-supervised learning based on self-representation subspace is an important topic that is closely integrated with reality.The paper focuses on self-representation subspace learning and regards semi-supervised classification or clustering as learning tasks.The research works are as follows:Based on the semi-supervised classification problem of self-representation subspace,we propose a latent subspace representation for semi-supervised classification.This method applies the self-representation based subspace learning to semi-supervised classification task.It introduces a filter matrix and builds a novel and effective model by combining the linear classification model with the learned a new subspace representation.Compared with the original features,the learned latent subspace representation is more discriminative.The proposed model jointly learns the subspace representation,latent space projection and classifier in a unified framework,and achieves a promising classification performance,which is verified by experimental results.Based on the semi-supervised clustering problem of self-representation subspace,this paper proposes a constrained tensorized multi-view subspace clustering.This method extends the case of single-view constrained subspace clustering to multi-view data.To exploit different views,the subspace representation matrices of different views are regarded as a low-rank tensor,which effectively models the high order correlations of multi-view data.To incorporate supervised information,a constraint matrix is devised to guide the subspace representation learning.Combining low-rank tensor and constraint matrix,we construct a flexible and efficient semi-supervised multi-view subspace clustering model.The experimental results demonstrate the promising clustering performance of our method.