A Recommendation System Fusing Multiple Relationships

There are many kinds of complex systems in the real world,and we use complex networks to describe their structures.In the multi-relational complex network model,subnets interact with each other,and each subnet has special topological properties,which determine the functions of complex systems.Recommendation system is an effective means to extract personalized information from mass information.In this paper,we improve the recommendation algorithm for a kind of multi-relational complex network,social network.By mining the network structure and improving the algorithm,we can predict the relationship between users and items more accurately.Recommendation system is an important method to process data and information in modern society.The recommendation algorithm based on matrix decomposition can effectively improve the performance of recommendation system.However,it is plagued by data sparsity and cold start problems.Incorporating the multivariate relation into the matrix decomposition model can increase the users and items' amount of information and solve the problems of data sparsity and cold start to a certain extent.Based on this,we fuse group information and multi-relational network information into the recommendation system which improves the performance of the recommendation system.In this paper,we add the group information to the matrix decomposition model and the multi-relation network information to the loss function.To improve the efficiency of decomposing the rate matrix,we use the algorithm of embeds Back-fitting into alternating least squares to get the latent features and use the least square method to estimate the regression relationship between subnets and target network.The performance of the algorithm is evaluated by using the public dataset of YouTube social network and Digg network.We analyzed the topology structure of the data,and divided YouTube users groups based on the number of user friends,and divided Digg story groups and user groups based on voting history and the number of user friends.To solve the problem that only implicit feedback data exists,sampling the negative cases based on Mean Variance Test.To YouTube dataset,we compared the recommendation performance with GSSVD and saltonbased recommendation algorithm;to Digg dataset,we compared with GSSVD and NBIbased recommendation algorithm.Under the standard of top-K,the method proposed in this paper has higher accuracy,recall rate and F1 than the other algorithms.