Research On Application Of Network Representation Learning In Recommendation Systems

With the informationization of society,the problem of information overload has become increasingly serious.The recommendation system can save users' time and improve the utilization of information.It examines characteristics of user,item,information and their historical information to form a judgment on users' preferences and items' characteristics,and then recommends items or information that meet the user's taste to the corresponding users.This thesis focuses on the rating prediction problem in the recommendation system.It introduces network representation learning,uses rating data,network data of trust between users and item category data,and the framework TrustEmbed and its extended versions to solve the rating prediction problem is constructed.The thesis mainly includes the following work:(1)In TrustEmbed,the trust network data and the user's rating data are merged into the Item Relation Network,and then the network representation learning method is used to obtain the low-dimensional item representations.The user's unknown ratings of items are then predicted by calculating the similarity between items.(2)The network representation learning method node2 vec referenced in TrustEmbed is improved,so that it can collect higher order topology information of global network.(3)The structures of the data set are analyzed from multiple perspectives.The prediction errors of every algorithm are evaluated.The influence of different parameters on the algorithm is discussed.T-SNE algorithm is used to visually analyze the item representations obtained by TrustEmbed,which proves that TrustEmbed can mine the potential features of items and provide effective help for the rating prediction task.(4)The sparseness of the dataset and its impact on each algorithm are explored and the effect of each algorithm in the absence of trust data is discussed.The results show that TrustEmbed has better robustness.Other kinds of network representation learning methods are used in TrustEmbed,which shows the flexibility and rationality of TrustEmbed.(5)Two extended versions of TrustEmbed are proposed to integrate the item genre data into the construction of Item Relation Network,which shows that TrustEmbed has strong extendibility.