Research On Recommender Systems Using Graph Neural Networks

The rapid development of Internet technology and the popularity of mobile terminals provide users with massive information resources,and users are gradually accustomed to online shopping,watching videos,listening to music,and browsing news.However,these rich resources also bring information overload,which leads users to need to spend a lot of time and energy to retrieve the content of interest from the mass of information.How to help users find what they need comprehensively and accurately has become the personalized recommender system's central goal.Recommending yet-unvisited points of interest(POIs),which may be of interest to users,is one of the fundamental applications in location-based social networks.Previous studies either develop matrix factorization-based approaches or utilize deep learning frameworks to learn better representations of users and POIs to estimate users' latent preferences.However,they have difficulties effectively using rich semantic information,such as social influence and geographical constraints.As a type of directed heterogeneous graph,the knowledge graph(KG)– where nodes and edges denote the entities and relations,respectively – is ideal for depicting side information associated with users/items.Recently,numerous studies have adopted KGs to enhance recommender systems.Despite their effectiveness,several crucial factors are missing in existing KG-based recommender systems.Firstly,previous studies mainly exploit only the user-item interaction links when finding similar neighbors for enhancing users' embeddings and usually treat each relation independently from all others.Besides,these methods often face inherent challenges in recommendation: data sparsity and cold start.This thesis designs two models,HGMAP and ERSPD,to solve the above issues.The main contributions of these two methods are as follows:(1)This thesis designs a new POI recommendation model called HGMAP,which leverages two GCNs to model the user social relationship and geographical distance,respectively.Moreover,HGMAP uses the multi-head attention mechanism to differentiate user preference over different aspects of POIs.(2)This thesis presents a KG-enhanced recommendation model named ERSPD,which computes the similarity degree of high-order relations and low-order relations to determine the importance of multi-hop neighbors on the computation of entity representation.Also,ERSPD exploits users' preference differences from users' rating records to optimize users' embeddings.(3)This thesis conducts extensive experiments on six real-world datasets,and the results demonstrate that the superiority and effectiveness of the two methods outperforming existing state-of-the-art baselines.