Learning To Recommend With Hidden Factor Models And Social Trust Ensemble

With the mass of information in the Internet,information overload is a problem of all Internet users already have or will soon encounter,and correlation is one of the main methods to solve the problem of information overload.Personalized recommendation on e-commerce site also highlights the increasingly important role.The technology is based on the discovery of the potential needs of users,filter out irrelevant information or items,and finally take the initiative to recommend to the user to meet their project needs, in order to achieve the purpose of ease information overload.Probabilistic matrix factorization model(PMF) based on the user-item rating matrix as the main and even unique data information,fuse the trust relationship between users and friends, linear polymerization the impact of friends on the user,to get the user’s potential characteristic vector which not only consider by the user’s own interest but also consider by the influence of the friend’s interest,to constraints on the potential features of the user obtained by the decomposition of the matrix.At the same time, a parameter is used to limit the extent of the constraint.And considering the text information of the project can also reflect the attributes of the project to a certain extent.Therefore, using LDA model to deal with the topic of the text information to get the topic distribution vector.Then the index function is used to map the potential feature vector of the project to the theme distribution vector to constraints the potential feature vectors of the project which obtained by the matrix decomposition,to get a product recommendation model which based on the user feedback information and the user’s trust degree is proposed.In experiment,the method of gradient descent is used to maximize the posterior probability.At the same time, it also optimizes the potential feature vector, the potential feature vector and the subject distribution vector,get the final theme distribution matrix, the user’s latent feature matrix and project potential characteristic matrix,finally get the user’s predictive rating.This theme carries out experiments on the real data of Yelp comment website.The mixed model was improved by 9.7% and 8.7% respectively in the typical measure of recommendation accuracy standard MAE and RMSE.Experimental results show that the results based on user feedback and user trust product recommendations model is superior to the recommended probability matrix algorithm and verified the effectiveness of the proposed hybrid model on the mitigation of the data sparsity problem.