Clustering Algorithms Based On Deep Neural Networks

Clustering is one of the most fundamental unsupervised learning problems in machine learning and its main goal is to gather similar data points in the data into the same cluster.Because of various redundant and complex structures of data in the original data space,clustering algorithms usually are difficult to separate different clusters from the general data,and the clustering effects are not obvious.In recent years,with the development of deep learning in various application fields,it can automatically extract feature representations that are abstract,nonlinear and more clustering-friendly from complex data structures for improving the performance of the algorithms.Therefore,we relies on the feature learning of the deep neural network to supplement the clustering algorithms so that the algorithms can capture data itself or its internal structures to better separate the clusters.Firstly,we study the theories of deep neural network,stacked auto-encoder and clustering analysis,analyze the differences between sparse auto-encoder,denoising auto-encoder and contractive auto-encoder,and introduce two clustering algorithms of self-organizing map network and Gaussian mixture model.Secondly,we propose a deep clustering algorithm based on self-organizing map network.It uses a stacked auto-encoder to learn the features of raw data,and then uses self-organizing map network to cluster the feature space in the unsupervised way.The data of feature space can be clustered into different data clusters by the model training and reach the maximum separation.In the experiment,we analyze the influence of each parameter on the unsupervised clustering accuracy of the data in the dimensionality reduction model and compare the difference of the image data before and after the dimensionality reduction.Analyze the effect of the relevant parameters on unsupervised clustering accuracy and compare the performance of the clustering accuracy between different algorithms.Finally,we propose a deep clustering algorithm based on Gaussian mixture model,which combines stacked auto-encoder and Gaussian mixture model to clustering the data.The proposed algorithm uses expectation maximization algorithm to train Gaussian mixture model and update data clusters after data dimen-sionality reduction,so that data can be better clustered in feature space.The effects of the related parameters of the algorithm on their clustering performance are analyzed.The performance of the algorithm on different datasets is analyzed and the clustering accuracy is com-pared with the different algorithms.In order to verify the effectiveness and validity of the proposed algorithms,experiments for the five datasets were done in the above algorithms.Through the experimental analysis and comparison,the proposed deep clustering algorithms improve the clustering performance on high and low dimension datasets,and verify the effectiveness of the proposed models.The deep clustering algorithms with feature learning show a better clustering effect and take less time in the clustering process.