Deep Subspace Clustering Based On Autoencoder

Clustering is widely used in many real-world applications.Subspace-based clustering is one of the most important methods.Recently,due to the powerful capability on feature extraction of deep network,extensive deep subspace clustering methods based on autoencoders have been proposed.However,these methods still have some shortcomings.For example,how to learn the more discriminative feature representations for better clustering results faces challenging.In addition,with the increasingly easy access to multi-view data in many areas,how to effectively utilize the consistency and diversity information contained in multi-view data simultaneously has become a hot research topic.To solve the above problems,this paper proposes a single-view deep subspace clustering method and a multi-view deep subspace clustering method based on autoencoders.The article includes the following two aspects of research:(1)The single-view deep subspace clustering method: this paper proposes a supervised feature learning module,which minimizes the KL divergence between the target probability distribution and the generated clustering probability distribution so that the model can learn more favorable features for clustering tasks.Meanwhile,for the mutually reinforcing relationship between feature extraction and data self-expression,the autoencoder module,the self-expression module and the supervised feature learning module are jointly optimized,which results in better clustering results.Experimental results on five real datasets sufficiently demonstrate the effectiveness of the model.(2)The multi-view deep subspace clustering method: the proposed method ensures that the network learns the general properties of all views and the specific properties of each view to make full use of the relationship between the multi-view data.Additionally,a diversity term is employed to view-specific matrices for encouraging the distinctness on the specific components.To effectively utilize the clustering results generated each iteration,an iterative self-supervision module is incorporated to promote the learning of the affinity matrix.Finally,the classical gradient descent algorithm is used to jointly optimize all the network modules,then spectral clustering is used to generate clustering results.We extend the linear model to the deep autoencoder structure for better exploiting the non-linear information contained in the multi-view data.Extensive experimental results on six real-world datasets demonstrate that the proposed model achieves uniform superiority over the benchmark methods and has good robustness.