Research On Enhanced Deep Subspace Clustering Algorithm Based On Contrastive Learning

In recent years,with the rapid development of the information age,a large amount of data has been generated on the Internet.If these data can be effectively used,it will greatly promote the progress of society.In order to fully mine the information in the data,cluster analysis is a very effective means.However,as the data becomes more and more complex,traditional clustering algorithms can no longer meet daily needs,coupled with the rapid development of deep learning,clustering algorithms based on deep learning emerged as the times require.Due to the excellent performance of contrastive learning in various learning tasks,this thesis considers applying contrastive learning to deep clustering tasks;specifically,the innovative work completed in this thesis is summarized in the following two aspects:1.This thesis improves the deep clustering via weighted k-subspace network and proposes a deep k-subspace clustering algorithm based on graph attention network.Specifically,based on deep clustering via weighted k-subspace network,this thesis enhances its feature extraction ability by introducing contrastive learning and graph attention mechanism;among which the deep feature learning module is trained based on contrastive learning to replace auto-encoder,and then the extracted features are aggregated by a graph attention module to further strengthen its feature extraction ability.In order to verify the effectiveness of the proposed algorithm,a large number of comparative experiments and ablation experiments were carried out on the MNIST and Fashion-MNIST datasets,and the experimental results verified the effectiveness of the proposed algorithm.2.This thesis improves neural manifold clustering embedding algorithm and proposes a self-enhanced deep subspace clustering algorithm.Specifically,this thesis uses the pseudo-labels generated by neural manifold clustering embedding algorithm to construct a selfenhanced mechanism,thereby designing a self-enhanced deep subspace clustering algorithm.Among which the feature learning module is trained based on the Maximum Coding Rate Reduction Criterion,and pseudolabels are generated based on the learned features,and then a specific sharpening operation is performed on the pseudo-labels to obtain more directional pseudo-labels.The loss function constructed by KL divergence(or square error)is used to reduce the distance between the two pseudolabels before and after sharpening,so as to realize the self-enhanced of the clustering results.In order to verify the effectiveness of the proposed algorithm,this thesis conducts comparative experiments on CIFAR-10,CIFAR-100 and STL-10 datasets,and conducts ablation experiments on each module in the model.The results verify that the proposed algorithm can significantly improve the clustering accuracy.