Research On Micro Blog Forwarding With Beta-Possion Factorization Based On Dirichlet Nonparametric Model

Social application has become indispensable in people's life.Microblog is one of the representatives.It breaks through the barrier of strangers' communication,so that hundreds of millions of people can communicate freely in the network.It is this feature that makes the data volume of microblog very enormous.The forwarding behaviour of microblog is an important resource to study user behavior in the microblog environment,Studying microblog forwarding is the best way to see user preferences,so it is meaningful.Users' preferences are influenced by many factors at the same time.This paper mainly studies the content and authors.Because this essay studies on the theme,we can judge whether the user's forwarding is affected by the author or the text content according to the subject of the microblog text,so that we can know more about the forwarding mechanism of microblog.At the same time,we can get the user's preferences.With the user's preferences,we can predict his future behavior.The research on group preferences has a wide range of applications,such as presidential election,to investigate the public opinion on the Internet.For example,it makes sense to judge whether to build a new shopping mall according to the preferences of residents in a certain area.There are many mature methods for the research of microblog forwarding,such as F-FFM method [1] and F-diff method [2].The former combines graph method and the latter is based on collaborative filtering method.The beta position model based on the Dirichlet nonparametric model in this essay has its own unique advantages in dealing with sparse non negative data,that is,to solve the problem of negative feedback.The nonparametric model introduced in this essay can adapt the number of subjects of new samples,so as to solve the problem of over fitting.Some traditional machine learning methods,such as randon forest and FM,have been studied a lot in solving this problem.Their main idea is to change the influence factors into eigenvectors,which are discriminant models.However,these methods also have limitations,such as the inability to solve the negative feedback problem and the sparsity of the forwarded data,as well as the limitations of the text vector.When they model the text,they often need to give some assumptions to the text,such as when they are characterized,their dimensions are all given,whether using LDA method or embedding method,this method It can fit the old data very well,but when the new text vector meets the new subject dimension,it cannot be explained very well.To solve these problems,this paper introduces a beta position model based on Dirichlet nonparametric model.Its advantage is(1)For the problem of negative feedback and data sparsity,a beta position decomposition model is proposed,which can only use the positive sample data,and the probability graph model effectively uses the sparse matrix.(2)Aiming at the problem that all parameters and their dimensions need to be calculated for each new sample,the method of stochastic variational inference is adopted.(3)In view of the problem that the number of topics set in advance is not enough due to the large amount of sample data,the Dirichlet nonparametric model is introduced,which can adapt the number of topics and make great use of the new sample data.(4)In order to solve the problem of whether users' tweets are influenced by content or author,the variable ? is introduced into the decomposition of position,which can be used as the weight of vector to observe users' forwarding preferences after all parameters converge.This paper discusses the research background and practical significance of Microblog forwarding,and then discusses some traditional methods and their shortcomings.Then in Chapter 2,it introduces probability graph model and its solution method,a model based on Gauss decomposition,and introduces the Dirichlet process and its Non-parametric characteristics,which will be very important in Chapter 3.Later,it will be discussed.The stochastic variational method is used to solve our probability graph model.In this paper,Micro blog forwarding data,which contains user forwarding records and Micro blog text content,as a data set.In the test set,we predict the probability of user forwarding,and evaluate the model with F1,NDCG and m AP as indicators,and carry out comparative experiments with F-FFM,F-Diff and IBPF as reference benchmarks.From the experimental results,we can see that the model used in this paper is 23%,2.6%,6% higher than the IBPF method,and 14.3%,7.3% and 2.1% higher than the F-Diff method.Moreover,by observing the variable ?,we can explain that the model in this paper is indeed affected by the topic.