Incremental Modeling Of Multi-dimensional Preference Based On Bavesian Network With Latent Variables

Massive rating data are generated by users' rating behaviors on Web2.0 applications such as Taobao,Jingdong,Amazon,Douban,Public Comment,Tripadvisor,etc.These data intuitively reflect users'evaluation of goods or services,and also contain potential user preference.Rating data not only relate to a wide range of applications,but also have the characteristics of sparseness and multidimensionality and so on.Generally speaking,the rating data include attributes of users and object evaluated by users,as well as rating value.Modeling with user preference based on rating data,we would be able to construct a kind of user preference model that can describe the dependencies between the variables of the rating data and be of great significance for the realization of personalized service and precise marketing.In addition,it is necessary to incrementally construct the model with rating data in order to reflect the dynamic evolution of massive rating data.Bayesian Network(BN),which is an effective framework for representing uncertain dependencies between attributes,has been widely used in preference modeling,but it cannot visually describe the implicit knowledge in data and has complicated dependencies between attributes.Latent variables are used to describe implicit knowledge in BN with latent variables.,which can intuitively describe users'potential preference and make the model interpreted easily.At the same time,BN with latent variables would be able to be constructed with EM algorithm and SEM algorithm.Considering the rating data,therefore,we use multiple latent variables to represent multi-dimensional preference and focuses on the method of construction and incremental update of Multi-dimensional Preference Bayesian Network(MPBN).Also,we overcome the efficiency bottleneck that computational complexity of constructing MPBN will rise exponentially with the increase of latent variables.Specifically,the main research contents in this thesis are summarized as follows:(1)Aiming at modeling of multi-dimensional user preference,BN with latent variables is used as a basic framework of representation and reasoning of knowledge.Further,the definition of a multi-dimensional preference model is given in this thesis.(2)For the dynamic evolution of massive rating data,the sensitivity of EM algorithm and SEM algorithm to initial values,and the efficiency of a large number of iterative calculations,we give constraint conditions and initial model,and then propose a method of construction and incremental update of MPBN based on the constraint conditions and nested merging of subgraphs.Furthermore,the parallel algorithm is implemented with Spark calculation engine.(3)According to the difference between individuals,an estimation method of preference and rating value is presented based on probabilistic reasoning of MPBN.This estimation method can correct estimation result with the introduction of users'rating behaviors.(4)In order to test the validity and feasibility of the method proposed in this paper,MovieLens dataset is used to test construction,incremental update,preference estimation and rating estimation of MPBN.Experimental results verify the effectiveness scalability and efficiency of our proposed algorithms.