Online Learning Method Of Latent Variable Model Oriented To User Preference Modeling

With the rapid development of mobile internet,Web 2.0 applications such as e-commerce and online social network are rapidly popularized,resulting in massive and dynamic user behavior data.User rating data is an important user behavior data,which reflects the user's preference for related products or services.By analyzing these user rating data and establishing user preference model,it can provide important data and technical basis for personalized services such as precise advertisement recommendation and user portrait,and it is of great significance.In recent years,many researchers have proposed user preference models based on latent variable model to describe the implicit knowledge in the rating data,thus simplifying the model and enhancing the interpretability of the model.As one of the latent variable models,Bayesian network with a latent variable(BNLV)can describe user preferences implicit in the rating data,and can describe any form of dependency and uncertainty among attributes in the rating data.Traditional latent variable modeling methods are usually based on machine learning algorithm of batch training mode,which assumes that all rating data could be used in the whole training process.However,data usually arrives in sequence,user preferences may change dynamically,and the number of rating data generated by users will gradually decrease with time.Because of the sparse and real-time rating data,the preference modeling algorithm based on batch training mode is difficult to adapt to online applications in the real world.In conclusion,the main contents of this thesis are summarized as follows:(1)The user preference model(UPM)based on BNLV is defined by analyzing the user preference and its related attributes implied in the rating data.(2)Aiming at the problem that the traditional BNLV parameter learning algor:ithm needs to iteratively calculate a large amount of data using Expectation Maximization(EM)algorithm,an online learning algorithm of UPM parameters is proposed by using Bayesian estimation extended Voting EM algorithm.(3)Aiming at the problem that EM algorithm is needed to learn parameters when classical Structure Expectation Maximization(SEM)algorithm learns BNLV structure,a method of reconstructing model parameter table when searching candidate model structure is proposed,and a scoring&search algorithm is improved to make UPM structure learning meet the needs of online learning.(4)Based on two real data sets,the efficiency and effectiveness of online learning of UPM parameters and structures are tested,and the related experimental results are analyzed.