Deep Clustering Algorithm Investigations Based On Matrix Factorization And Autoencoder

Unveiling the intrinsic information of the data in depth is crucial when pursuing the performance of a learning strategy,as the deeper hierarchical features could provide extra essential and discriminative information concealed in the data.At present,several technologies have been involved with this challenge,such as deep matrix factorization and deep learning.Clustering is a fundamental data analysis task that highly depends on the feature qualities to partition objects into different groups.Clustering algorithms armed with deep representation learning abilities are referred to as deep clustering(DC)models and have continually shown their advantages compared with plain vanilla ones.Based on these observations,this dissertation dives into the DC research and comes up with two main innovations detailed in the following contents.Recent studies have suggested that the performance of the clustering models built with the deep matrix factorization procedure surpasses the single-layer formed ones in consequence of perceiving the hierarchical information.However,the up-to-date clustering algorithms fail to simultaneously integrate Multi-View Clustering(MVC)and deep Concept Factorization(CF)into a unified framework.In this dissertation,a novel DC model is proposed to tackle this challenge,which brings deep CF to MVC for learning the hierarchical information through performing multi-layer CF and derives a common consensus representation matrix to fetch the shared features among different views.Besides,we impose manifold regularization on each view for locally geometrical structure retention and employ Gaussian kernel to map the original space to a higher Hilbert space for effectively distinguishing the data points.Finally,an efficient optimization algorithm with theoretically guaranteed convergence is developed to solve the proposed model.Experiment results on several open datasets demonstrate the superior performance of the proposed model compared with baseline methods.Embracing the deep learning techniques for representation learning in clustering research has yielded stunning results over the last few years,ushering in a new era of DC exploration.Typically,these DC models capitalize on autoencoders to learn the intrinsic features that facilitate the clustering process in consequence.Nevertheless,as the plain Variational Auto Encoder(VAE)is insufficient to perceive the comprehensive latent features,which could lead to deteriorated clustering results,this dissertation designs a new DC method to address this issue.Specifically,the generative adversarial network and VAE are coalesced into a new autoencoder called Fusion Auto Encoder(FAE)for discerning more discriminative representation that benefits the downstream clustering task.Besides,the FAE is implemented with the deep residual network architecture which further enhances the representation learning ability.Finally,the latent space of the FAE is transformed to an embedding space shaped by a deep dense neural network for pulling away different clusters from each other and collapsing data points within individual clusters.Experiments conducted on several image datasets manifest the effectiveness of the proposed DC model against the state-of-the-art methods.At last,this dissertation investigates the clustering performance of the FAE model when imposing the manifold regularization on its mean vector latent space.And the experiment results imply that adding the regularization can actually boost the clustering accuracy.