Research On Recommendation Model Of Multi-source Information Fusion

With the rapid development of social media and mobile applications,the problem of information overload is becoming more and more serious,and it is becoming more and more difficult for people to obtain useful information from massive amounts of data.The generation of recommendation systems can effectively solve the problem of information overload and provide users with personalized recommendation services.In recent years,they have received extensive attention from researchers.Collaborative filtering is currently the most widely used recommendation algorithm.However,the collaborative filtering algorithm only uses the interaction data between the user and the project.When encountering data sparseness and cold start problems,the recommendation quality is not high and cannot provide users with satisfactory recommendations.In response to these problems,researchers have proposed a large number of solutions.However,just using the score data can not fundamentally solve the problems faced by the recommendation system.The information generated by Internet users is becoming increasingly abundant,providing multi-sourced data,such as project attribute information,social network information,geographic location information,and user comments.These independent data sources provide opportunities for solving inherent problems and improving the performance of the recommended system.Based on the existing work,this paper studies the recommendation algorithm for fusing multi-source information.The main research contents and research results of this article are summarized as follows:(1)A matrix decomposition recommendation method So Reg IM for social trust network reconstruction is proposed.In view of the fact that the current social network data is mainly binary data(edge 1 and edge 0),a community discovery algorithm based on signal propagation is used to calculate the topological similarity between users in the social network as the user's trust value Size and reconstruct the social trust network by setting thresholds.The new social trust network not only contains user trust,but also effectively uses the indirect neighbor information in the original social network.Then,we merge the reconstructed social trust network with the matrix decomposition model,use the user score and the user's social trust relationship to calculate the user's similarity so as to constrain the decomposition form,and optimize the learned user feature vectors and items Represents a vector.The experimental results on the classic data set show that our method is superior to existing methods compared with similar social recommendation algorithms.(2)A recommendation algorithm RENFAR that integrates user-project community and comment information is proposed.The neighbor information in the social network has a great effect on the performance of the recommendation system,however,the social information only exists in specific scenarios.This article uses a community discovery algorithm to find user and project community information in the user-project interaction matrix,and interactively encode community neighbors(including users and projects)through convolutional neural networks.At the same time,we use sentiment dictionaries to preprocess the review text,and convolutional neural networks to encode the relevant review text.Then,the neighborhood feature and comment feature of the community are fused through feature combination.Experimental results on commonly used public data sets show that the RENFAR model we have proposed has greatly improved the recommended performance.