Research On Heterogeneous Graph Neural Network And Its Application In Recommendation System

Recommender systems are present in the mainstream applications of many companies both at home and abroad.Recommender systems can not only facilitate user decision making and improve user experience,but also increase user retention rate to enhance business profits.In the field of recommender systems,most of the data have(heterogeneous)graph structure.In the past few years,heterogeneous graph neural networks have shined in the field of recommendation systems.Hence,it is very practical and valuable to apply heterogeneous graph neural networks to recommender systems.Taking the application of heterogeneous graph neural network in recommendation system as the research object,two social recommendation models CFGRec and ConsistCFGRec based on heterogeneous graph neural networks were proposed,and a social food recommendation system was designed and implemented in this thesis.Existing studies usually directly perform multi-layer aggregations on the neighbors of user and item nodes,ignoring the idea of collaborative filtering that similar nodes have gaining effect on the target node representation learning.Therefore,in this thesis,a collaborative filtering enhanced heterogeneous graph neural network CFGRec for social recommendation was proposed.According to the idea of collaborative filtering,this model constructs a similarity graph between users(items),and controls the number of edges in the graph by adjusting the similarity threshold;The representations of users(items)in similar space,social space and user item interaction space are dynamically and adaptively fused through attention mechanism.Existing studies usually ignore the social inconsistency problem which is common in social recommendation,and the user-item interaction bipartite graph tends to be sparse.In this regard,in this thesis,a heterogeneous graph neural network Consist-CFGRec which is applicable to alleviate the social inconsistency problem was proposed.ConsistCFGRec,in order to alleviate the social inconsistency problem,adds a multi-headed selfattention mechanism to capture consistent friend nodes in the message passing between user nodes in the social relationship graph;to capture the differences in attractiveness of different items to target user,and the differences in preferences of different users to target item,adds a multi-headed self-attention mechanism to the message passing between user nodes and item nodes in the user-item interaction bipartite graph.Experiments on three representative public social recommendation datasets Film Trust,Ciao and Epinions,show that CFGRec has 2.56%,2.55%,1.25%,1.51%,1.24%,and 1.38% recommendation performance improvement in RMSE and MAE,respectively,compared to the existing baseline model.Consist-CFGRec also has 2.85%,2.98%,1.84%,2.20%,1.44%,and 1.91% recommendation performance improvement.The implemented social food recommendation system application in this thesis completely implements the basic functions of industrial social recommender systems.Users can view the recommended food,establish social relationships with friends and share daily,etc.Administrators can manage models,users,food and ratings in the system management background.It helps to understand the recommendation models proposed in this thesis.