Research On Clustering Methods Based On Deep Learning

In all kinds of machine learning algorithms,the clustering algorithm is one of the important parts.With the popularization of the internet and the application of computer technology,the increasing network activities make the amount of data grow exponentially.With the emergence of huge unlabeled data to be processed,the application of clustering algorithms is increasingly widespread and many effective clustering methods are proposed.The key of clustering algorithms is to preserve the sample features and spatial relationship between the original samples in the clustering process.Shallow clustering methods use linear projection to reduce feature dimension and extract sample features,which usually have a weak feature representation ability.In contrast,deep clustering methods usually use nonlinear feature extraction methods,which have advantages in preserving the original sample features.However,most of the deep clustering methods pay too much attention to preserving the feature information within the samples but ignore the relations between samples,which limits the performance of clustering methods.Besides,a large number of neural network parameters needs to be optimized in the unsupervised deep clustering method,but the lack of label information will limit the representation ability of the deep neural network.Based on the existing deep clustering methods,this paper focuses on how to introduce the relationship between samples into clustering methods.By introducing the relationship between samples,the original spatial relationship between samples is preserved in the low-dimensional embedding space.In this paper,a deep clustering method based on the structured graph of adaptive allocation of adjacency samples is proposed.Then,a clustering method based on a generative adversarial network is proposed,which is guided by Gaussian distribution fitting similarity matrix.The main contributions of this paper are shown as follows:(1)A structured graph based deep clustering network is proposed to simultaneously perform deep feature representation learning,structured graph learning,and clustering.The deep embedding processing can preserve the manifold structure of samples more accurately to pursue an ideal clustering structure.This model also performs feature learning by optimizing the loss function of KL divergence based on the clustering objective with a self-training target distribution.In this network,deep feature learning,structured graph learning,as well as data clustering,are jointly optimized and they can enhance each other.(2)A semantic structure based deep adversarial clustering model is proposed to learn a better feature extraction model and can retain the original features of samples better.The relations between the distance of samples and the similarity probability of samples are also analyzed in this model.According to the analysis in the paper,by using two semi-Gaussian distributions to fit the sample distribution,we generate a structured graph that can effectively preserve the similarity relationship between samples.The feature extraction model,similarity matrix between samples,and the clustering method are optimized together to obtain high-quality clustering results.