The Personalized Recommendation Algorithm Based On Collaborative Filtering

With the arrival of the internet plus and artificial intelligence era,it is difficult for us to quickly get information which we really need in daily life from the mass information.In order to solve the problem of information overload,recommender systems came into being,which aim to offer users accurate and personalized product or information services,so as to improve the user's experience.Among them,the collaborative filtering algorithm is one of the most popular algorithms investigated in academia and industry.However,as the number of users and items in the recommendation system increases exponentially,the problem of data sparseness seriously affects the performance of the recommendation system.In this thesis,basing on the research of existing algorithms,we analyses the shortcoming of the existing collaborative filtering algorithm.This paper combined the project content information,social relationship network and collaborative filtering algorithm to provide project recommendations for users.We proposed three novel collaborative filtering algorithms based on the collaborative topic model(CTR):user social relation incorporated into collaborative topic regression,collaborative topic regression with social regularization and collaborative deep learning with social regularization.This paper selected the open data set Last.fm to prove the validity and reliability of the proposed algorithm.The significance and innovation of this paper mainly reflect as follow:First,we proposed a model called User Social Relation incorporated into Collaborative Topic Regression(USRCTR)according to the basis of CTR,and combine all three-the user and item feedback information,item content information,and social relation network to construct a recommendation engine based on the novel hierarchical Bayesian model.Assuming that there is come connection between the similarity of users preference and the existence of users,we used the probability link function to evaluate the effect on the users'preference brought by the social relationship network,and the effect can be used as constraint of the object function.Second,we proposed a model called Collaborative Topic Regression with Social Regularization(CTR-SR),which extends User Social Relation incorporated into Collaborative Topic Regression by integrating the topic model,social relation work,and Bayesian probability matrix decomposition.We can capture the potential social relationship among users by using Laplacian regularization,and then directly embed it to the potential feature of users interest so as to explore the effect of social networks on users interest better.Third,we proposed a model called Collaborative Deep Learning with Social Regularization(CDL-SR),which extends Collaborative Topic Regression with Social Regularization by integrating Bayesians Stacked Denoising AutoEncoder(Bayesians SDAE),social relationship work and Bayesian probability matrix decomposition,and we construct a recommendation engine of deep learning based on the Bayesians model.Through deep learning,Bayesian SDAE can extract the effective feature representations in depth from the project content,and capture the similarity relationship among projects,so as to obtain the better potential feature representation.Fourth,experiments on the Last.fm dataset show that this model can achieve better prediction accuracy than several improved recommendation systems such as CTR,LACTR,CTRSMF,CTRSTE and provide users with diverse personalities recommendations by a series of parameter learning and optimization.