Modeling Dynamic Behaviors Of Users Based On The Probabilistic Graphical Model

With the rapid development of Web2.0 and the e-commerce applications,large-scale online rating data are generated,which makes it possible to analyze users behavior data and model user behaviors.Considering the dynamic property of the rating data and user behaviors,in this thesis we adopted Bayesian network with a latent variable(abbreviated as latent variable model)as the framework for describing mutual dependencies and corresponding uncertainties,and then constructed the model that can reflect not only the uncertainty of dependence relationships among the attributes in rating data but also the dynamic property of user behaviors,and further study the method incremental modifications for rating behaviors model.To obtain the dynamic rating behavior model from user's rating data,we first adopted the Bayesian Information Criterion(BIC)as the coincidence measure between the candidate model and rating data,and then proposed the scoring-and-search based method to construct the latent variable model that reflects user's rating behaviors.Then,we gave the method for filling latent variable values based on the Expectation Maximization(EM)algorithm.Further,we proposed the method for constructing the latent variable model between adjacent time slices based on conditional mutual information and irreversibility of time series.Finally,we employed the Maximum Likelihood Estimate(MLE)to estimate the CPTs for each node of the dynamic behavior model.To impliment the incremental modifications of the dynamic rating behavior model from new rating data,we first gave the definition of influence degree on the each node of current rating behavior model,and then we proposed the method for getting modify substructure based on the definition of influence degree and Markov Blanket(MB).Then we proposed a method for modify the directed edges between adjacent time slice variables based on conditional mutual information and time series irreversibility.Further,we adopted the BIC score as the coincidence measure of the candidate rating bavaior model,and proposed the method of constructing the directional edge between the variables in the time slice from the new rating data,which based on scoring-and search algorithm.The experiments which were established on the MobieLens data sets tested the running time of each algorithm in the process of dynamic evaluation behavior and the running time of the incremental modification algorithm.Meanwhile,we also tested the dynamic evaluation behavior model in this thesis,predicted the mean absolute error(MAE),accuracy,coverage and F1 values of user evaluation behavior.The experimental results verified the efficiency and effectiveness of the dynamic user behavior modeling method proposed in this thesis.