Clustering Analysis Based On Variational Auto-encoder And Mixture Model

The slogan of happy life and more convenience is constantly appearing in people's life nowadays.People's daily life more and more convenient,but also constantly produce a variety of data.This data is then aggregated into a sea of data that reflects the real world.In order to give full play to its value,we need available and effective data mining technology for data analysis,clustering analysis is one of them.Whether it is hidden in the data of commercial value or academic research value,can be used for different clustering algorithm mining.At present,there are many traditional clustering algorithms which can meet the requirements of the problem and the user's needs if the problem is analyzed in the direct and considerable low-dimensional data space.However,in the application of many real scenes,the collected data gradually show the characteristics of high-dimensional and non-Gaussian data.Therefore,the existing traditional clustering methods can not meet the requirements of clustering analysis.The appearance of Deep Learning brings a solution to this problem,which can find hidden structures in complex data and push the model to update parameters automatically to optimize the model.In this paper,we combine the unsupervised generation model with the mixture model in Deep Learning for clustering research,not only to achieve the automatic extraction of data features,but also to a certain extent to avoid dimension disaster.In this paper,we embed the mixture model into the Variational Auto-Encoder framework and propose two unsupervised clustering methods:(1)Unsupervised image classification based on Variational Auto-Encoder and Student's-T Mixture Model.This method aims at traditional clustering algorithm based on Variational Auto-Encoder,the network feature loss is too large to extract potential,hierarchical feature representation,and the clustering algorithm based on Gaussian Mixture Model is sensitive to outliers and can not describe the data with heavy-tailed featuresaccurately.In order to solve these problems,we improve the traditional Variational Auto-Encoder network structure,combine the Convolutional Neural Network to realize the hierarchical feature extraction,and construct a mixture model based on Variational Auto-Encoder using Student's-T Distribution for clustering;(2)Unsupervised clustering analysis based on Variational Auto-Encoder and Dirichlet Mixture Model.In the course of research,we found a clustering algorithm based on Gaussian Mixture Model and(1)Student's-T Mixture Model,it is not suitable for the description and fitting of bounded data,and the multiple parameter technique based on Gaussian Distribution family used in the original Variational Auto-Encoder is not suitable for the Dirichlet Distribution,we use flexible and easy-to-use Dirichlet Distribution to construct a mixture model based on Variational Auto-Encoder to meet the needs of bounded data fitting.For the Dirichlet Distribution used,we apply a new Reparameterization Trick that allows the algorithm to meet the variational requirements to produce a model that is usable and valid.