Estimating News Event Inducement Based On Bayesian Network With A Latent Variable

The continuous development of the Internet,search engines,and news platforms has produced massive scale of news event data.As an important type of data in the field of data analysis,news event data contains various valuable information.Modeling and estimating event inducements from news event data is significant for solving public opinion control,precise decision support,and user behavior orientation,etc.In addition,data w.r.t.news event has the characteristics of dynamics.Therefore,studying the incremental learning of the event inducement model is of great significance for accurately estimating the event inducement.As an effective framework for uncertainty knowledge representation and probabilistic reasoning,Bayesian Network(BN)is widely used in the analysis and modeling of data knowledge discovery problems.However,BN cannot intuitively and effectively describe the implicit knowledge in event data.Latent Variable Model is constructed by introducing latent variables into BN,which can describe and express the implicit knowledge in the data,and make the uncertainty dependence between explicit variables easily.Considering the data of event features,therefore,we use Bayesian Network with a latent variable as the basic framework for model construction and study the construction of event inducement model and incremental learning,and estmate event inducement based on probabilistic reasoning of event inducement model.In conclusion,the main contents in this paper are summarized as follows:(1)Firstly,the definition of Event Inducement Bayesian Network(EIBN)is given based on Bayesian Network with a latent variable.Secondly,aiming at the sensitivity of the initial value of the SEM algorithm,we give the structural constraints and initial structure of the model learning based on the event domain knowledge.Then,the size of the penalty is adjusted by taking the expected value of the BIC score and AIC score,and proposes the BAIC(Bayesian-Akaike Information Criterion)scoring method.Finally,an EIBN model construction method is given based on constraints,BAIC score and SEM algorithm.(2)In order to reflect the dynamic changes of news event data over time,we propose incremental learning method of EIBN model based on Voting EM algorithm,SEM algorithm and structural constraints,which is used to incrementally learn the established event inducement model and refit the model to the news event data.(3)In order to perform probabilistic reasoning based on the variable elimination method,an effective EIBN model is given.Aiming at the multiple possible values of the event inducement,a branch and bound algorithm is given to extract the optimal subset of inducement.(4)We verify that the validity and efficiency of EIBN model construction,incremental learning,and inducement estimation by experiments built on real datasets.