Research On Deep Subspace Clustering Algorithms

Traditional clustering algorithms cannot effectively process high-dimensional data and have high computational complexity.Based on these two problems,a subspace clustering algorithm is proposed.However,the effect of subspace clustering algorithms on nonlinear data is limited,although Kernel techniques are introduced,there is not enough reason to determine that the implicit feature space corresponding to the kernel is suitable for subspace clustering.With the development of deep learning,the deep neural network-based clustering method has a high representation ability,can be able to handle features effectively,this paper studies the deep subspace clustering algorithm.The specific research contents are as follows:(1)Deep subspace clustering algorithm based on denoising autoencoder.We introduce a denoising self-encoder to make the learned representation more robust,learn the latent space through the non-linear transformation of the network layered stack,and use the self-representation layer to learn the similarity matrix for subspace clustering in the latent space.Then use spectral clustering to complete the clustering.The proposed method has better generalization performance due to its non-linear representation ability,and is especially suitable for high-dimensional data with significant correlation.Experimental results show that the model is effective for subspace clustering.(2)Improved deep subspace clustering algorithm.This method is based on deep subspace clustering network(DSC),and improves the problem of too large feature loss that affects the clustering result,and effectively improves the problem of feature extraction.The model adds a fully connected layer as the downsampling layer and the upsampling layer in the encoder and decoder parts of the convolutional self-encoder,so that it can further integrate effective features and learn feature representations more effectively.We compared five classic algorithms(including deep subspace clustering algorithms based on denoising autoencoders)on four data sets.Experimental results prove that the proposed model structure has made significant progress compared to DSC models.Clustering The effect is also superior to other classic unsupervised subspace clustering techniques.This paper has 10 figures 4 tables,and 99 references.