Research On Multi-view Mixtures Of Gaussian Process Models For Classifications

The Gaussian process is a powerful Bayesian nonparametric model.It can be used for both classification tasks and regression tasks.It has a wide range of applications in many fields.The proposed mixtures of Gaussain process model(MGP)has further improved its data modeling capabilities.However,the single-view MGP model cannot reasonably and efficiently model the multi-view data widely existing in reality,which limits the practicality of the MGP model.In order to improve the modeling ability of MGPs,especially the ability to model multi-view data,a new framework of multiview learning for the MGPs named the multi-view MGP model(MvMGP)is proposed and is instantiated for classification tasks in this paper.The specific improvements are reflected in three aspects.Firstly,the idea of co-regularization is used to design an objective function for the model based on two ideas: one is that in order to leverage the information of a single view,the posterior distribution of latent variables is maximized in each view;the second is that in order to maintain the consistency between views,the posterior distribution of latent functions in different views are constrained to keep as close as possible.Secondly,since the derivation of the posterior distribution involves complex integral operations and cannot be solved analytically in the process of solving model,two Bayesian inference methods are adopted to solve the model,namely the deterministic approximation method,variational inference,and the stochastic approximation method,Markov chain Monte Carlo(MCMC).Finally,in order to sample the conditional posterior distribution of the MvMGP model more accurately and efficiently,a new MCMC method named the decomposed slice sampling is proposed,and the correctness of this method is verified from both theoretical and experimental aspects.Eexperimental results on multiple real-world datasets have shown that the performance of the newly proposed MvMGP model is superior to the original MGP model and several state-of-the-art multi-view learning methods.In addition,the MvMGP model based on MCMC has higher accuracy than the MvMGP model based on variational inference.