Structured Auto-encoder Based On Deep Clustering Algorithm Analysis

Most of the existing deep clustering methods often use the minimized reconstruction loss to capture the high-dimensional probability distribution of the input,but the discriminative ability of deep features is not necessarily related to the reconstruction loss,and these deep clustering methods usually only focus on the attribute information of the learning samples themselves,but rarely consider the structural information between the samples.In order to solve the above problems,effectively improve the discriminative ability of deep features,and make full use of the structural information between unlabeled samples,and jointly optimize the feature extraction and clustering process of samples,this paper proposes a structured deep discriminant embedded coding network clustering algorithm(Structured Deep Discriminant Embedded Coding Network Clustering Algorithm,SDDECC).Firstly,a multi-layer convolutional autoencoder network is introduced to maximize mutual information and minimize prior distribution constraints to improve the ability to discriminate deep features and make the coding space more regular.Then,the transfer operator is used to integrate the features learned by the deep discriminative embedded coding network sub-module into the graph convolutional neural network sub-module to realize the collaborative learning of the sample's own attribute information and structural information,and effectively improving the feature discrimination ability while retaining more available structural information.Finally,using KL divergence combined with the potential feature space distribution generated by the dual network structure,a joint optimization framework that integrates feature learning and clustering is established,and end-to-end guidance is given to the network model to learn deeper features that are more conducive to clustering and to update parameters iteratively.The clustering accuracy of the SDDECC algorithm on the four classic image datasets of USPS,MNIST,Fashion-MNIST and STL-10 reached 79.86%,90.22%,61.71% and 38.65%,respectively.Compared with the sub-optimal clustering algorithm,the accuracy has been improved by 1.97%,2.16%,3.18% and 1.45% respectively.The experimental results show that the SDDECC algorithm produces a stronger prediction effect than other advanced algorithms,and better improves the clustering performance.There are 25 figures,11 tables and 64 references in this paper.