Research On Social Recommender System Based On User Context And Interaction Activities

In order to relieve the data sparsity and cold start problems of traditional recommender systems,social recommender system was proposed in this context.With the popularity of social communication platforms,users' social information contains a wealth of mining value.Thus,users' social information as an input element is widely concerned.The primary process of social recommender system includes two parts: constructing the trust network and generating recommendations.A qualified trust network can fully reflect the relationship between users and relieve the problems of data sparsity and cold start.In this thesis,we proposed “A Trust Network Constructing Model Based On Users' Interactions”(Soc_Pro).The model makes the advantage of users' interactions and trust propagation method to compute the global trust value of each user.Then,we introduced the time decay factor,and obtain the direct trust degree and similar trust degree by the interactions and similarity between users.This model overcome the problems of “Trust Network Based on Behaviors”(Soc_Trust),such as it just predict the relation between users but not trust degree,and it treat all the historical information equally in the prediction process.Then we choose the good interaction features by Greedy Feature Selection Algorithm of Matrices Factorization to avoid producing new noises during the latent feature vectors learning period of Factorization Machines.The method tackled the problem that traditional Factorization Machines assumed all features to be interacted with all other features but some of them are meaningless.At last,we evaluate our proposed model Soc_Pro and compared with Soc_Trust on accuracy of trust prediction,it demonstrated that our model performed better than Soc_Trust.The recommendation experiment demonstrated that our recommendation generation model CI_FM performed better than Factorization Machines,Matrices Factorization on accuracy of recommendation item.