The Study Of Multi-view Subspace Clustering Modeling And Its Applications

Real-world data often have high ambient dimension and complex structure,which brings difficulties to processing of data.However,high-dimensional data sets are often inherent of low dimension.In fact,high-dimensional data usually fall in the union some low-dimensional subspaces.Therefore,high-dimensional data can be clustered by using the representations of subspaces based on spectral clustering.In addition,many of the problems in computer vision involve multiple feature sets,where different feature sets represent different information about the data set,for example,images and videos can be described by color,texture,and edge,etc.Therefore,multi-view subspace clustering has emerged.Since different views are usually collected from different domains by using different feature extractors,they may contain information that is complementary to other perspectives.A key issue in multi-view clustering is how to better utilize the information from multiple views.In this paper,two works have been done.First,an ₁l-norm regularization model is proposed based on the multi-view complementary information.It utilizes the idea of voting in integrated learning and obtains a unified self-representation matrix by using the voting method.And this model has combined spectral clustering and self-representation into a unified framework,which is called the ₁l-norm regularized multi-view subspace clustering model.Second,the traditional similarity based on space distance is a kind of local similarity while the self-representation coefficient is global similarity.Using the local similarity guide the global similarity,a weighted sparse regularized representation model is proposed.Combining the weighted sparse regularized representation model with the ₁l-norm regularized multi-view subspace clustering model,a weighted sparse ₁l-norm regularized multi-view subspace clustering model is proposed.It not only utilizes the complementary information between the views and the consistency of the labels,but also considers the local and global similarity of the data.Experiments have been carried out on the face data sets commonly used in several cluster analysis.Experimented results show that the proposed method has better clustering performance than other related methods.