Researching On Deep Bayesian Topic Model

Poisson Gamma Belief Network(PGBN)is a deep topic model based on bayesian framework,which can extract the multi-layer characteristic information of data through multi-layer nonlinear network,and have good classification performance in text and image data,but there are few works about application analysis of PGBN model,so the practicality study of the PGBN deep model is of great significance,this thesis researches a web page documents real-time processing application.When PGBN model was proposed,the parameter's inference method is gibbs,and the training method is batch learning,our group proposed an online learning method based on the PGBN model,this method using stochastic gradient method to update global parameters and gibbs method for other parameters,which randomly selecting minibatch to train model has some advantages in the convergence speed of model training.But the classification performance,characteristics and application ability are unknown.So,it is of great significance to study the PGBN model based on the online learning method.This paper use the PGBN model which based on gibbs inference method as baseline,researches the classification and characteristics of the PGBN model based on online learning method,and analyzed the applications of PGBN model based on two kind of method by programming implementation.The main contents of this paper are as follows:In order to study the classification performance and characteristics of PGBN model based on online learning method,this papre firstly study deeply of the PGBN model based on Gibbs method,and obtaining the baseline by simulating the classification experiment of text and image data.And then,the PGBN model based on the online learning method is studied deeply,the simulation of text and image data classification based on the same model parameters is carried out,and the simulation results are analyzed and discussed at last.The PGBN model based on the Gibbs method and the PGBN model based on the online learning method are programmed in three different programming languages: C language,MATLAB language and Python language,in which C language has certain advantages in the efficiency of implementation,MATLAB has high matrix computing efficiency,Python has a lot of powerful and simple open source library,which provides a great possibility forthe application and extension of the model.In the programming simulation experiment,comparing and analyzing the time and space complexity of the PGBN model based on two different training methods.At last,analysing and discussing the feasibility of Python language realize the web page documents real-time processing application.Based on the PGBN model which based on the online learning method,this thesis research and analysis the web document real-time processing application which is realized by Python language.In this application,the crawler technology is used to crawl the web page document data,and the redundant information is deleted by regular expression to get the document data.Finally,using the C extern interface of python to implement the complex module of model algorithm to enhance the training efficency and practicality of model.