A Dynamic Topic Model On Evolution Of Social Network Users' Interest

In recent years,with the vigorous development of online social platforms such as Weibo and Twitter,users generate huge amounts of data on social platforms every day,of which text data occupies the majority.User data is of great significance for public opinion analysis,user portrait and personalized business recommendation.For such data,it is an effective way to extract text information through topic models.Topic model can make a better topic mining for the long static text,but the feature of data sparsity and the dynamic change of topics in the short text stream of the online social platform leads to the poor effect of the classic topic model in this scenario.Based on the research of dynamic topic model,this paper proposes a user-level topic model for short text in social platform.At the same time,in view of the deficiency in the model,the word weight is used to filter words with high frequency but low topic contribution in short text,and the dynamic theme model,which does not need to specify the number of topics,is realized by the hierarchical dirichlet process.The main work of this paper is as follows:1)The users' topics are expressed as the multinomial distribution of hybrid subjects,and the distribution of user and topics is modeled according to the user's text in current time slice and the distribution of history time slices,in order to infer user's current topic.Through the word pairs in short text to overcome the problem of the sparsity of words in short texts,the dirichlet generation model and its Gibbs sampling process are improved.Experimental results show that the dynamic social user topic model is promoted in the topic generalization performance and achieves more accurate topic distribution inference.2)In order to solve the problem of high frequency and low topic contribution of some words,the weight of words in text is calculated and the sampling algorithm of word pairs is improved according to the weight of words.Experimental results show that the introduction of word weight improves the quality of word sampling and the effect in topic prediction and user topic classification compared with the original model.3)In order to solve the problem that specified number of topics probably result in inaccurate topic identification,hierarchical dirichlet process is introduced for the initial model and a dynamic topic model of three-layer dirichlet distribution is established to realize the dynamic change of topic number in different time slices.Experimental results show that compared with the original model,the proposed model can dynamically adjust the number of optimal topics according to the time changes,and achieve more accurate prediction of user topics at different times.