Research On Recommendation System Based On User Behavior And Social Relations

In the Internet era,the amount of data that people are exposed to has far exceeded their own data processing capacity.As an efficient technical means to solve this problem,recommendation system has been studied more and more.In recent years,more and more researches began to apply graph neural network to recommendation system.However,there are still some problems in the existing recommendation algorithms based on GNNs.Firstly,the existing recommendation algorithms based on GNNs often do not pay attention to distinguish the importance of the neighbor nodes to the central node when calculating the embedding of the central node,and do not pay attention to distinguish the user's rating value for the item.Secondly,when computing the node embedding,the existing recommendation algorithms based on GNNs often only transmit the information of its first-order neighbor nodes while ignoring the information of its higher-order neighbor nodes.Finally,most of the existing recommendation algorithms based on GNNs only utilize the user's social network graph and the bipartite graph of the interaction between users and items,while ignoring the connections between items.Aiming at these problems,this thesis proposed new solutions and conducted experiments on relevant datasets to verify the effectiveness of the proposed methods.The main contents of this thesis are as follows:1.A rating-aware deep GNN model DGAT(Deep Graph Attention Network)is proposed.When computing node embedding,the model strictly distinguishes the rating values between users and items,and uses a GAT to distinguish the importance of different neighbor nodes.The model aggregates the high-order neighbor node information of the central node by stacking multi-layer GATs.Experimental results on three public datasets demonstrate the effectiveness of DGAT,which reduces MAE and RMSE by up to 2.74%and 2.98% compared with the optimal baseline model.2.A deep GNN model DGAT++ which can capture the item relationships is proposed.This model builds an item relationship graph for items.The model learns embedding for user nodes and item nodes from the user social graph,the user-item interaction bipartite graph,and the item relationship graph.Compared with DGAT,DGAT++ reduces MAE by up to 1.77% and RMSE by up to 0.74% on the three datasets3.Based on the recommendation model proposed in this thesis,a recommendation system based on user behavior and social relationship is designed and developed to implement store recommendation for users.