Research And Implementation Of Relation Aware Recommender System Based On Knowledge Graph

The rapid development of the internet pushes forward the development of society,bringing convenience to our life at the same time.Nevertheless,despite the benefits it brings,it also leads to the information overload problem.To address the abovementioned issue,recommender system emerges as the times require.Among various recommendation techniques,collaborative filtering based methods have achieved significant success,nevertheless,they suffer from the defect of unable to model side information.Traditional supervised learning based methods such as factorization machine are able to integrate side information such as item attributes and user profiles,however,they model each interaction independently and overlook the relation between instances as well as items.Integrating user-item interaction with item knowledge graph makes it possible to capture the collective behavior of users at the same time takes full advantage of information such as item attributes.Recently,incorporating knowledge graphs into recommender systems as auxiliary information has attracted much attention due to its rich semantic content.Knowledge graph is a heterogenous network which stores all kinds of information in the form of graph,in which a node stands for an entity and an edge represents relation between two entities.By incorporating user-item interaction graph with item knowledge graph we can consider users' collective behavior at the same time explore latent relations between items through their connectivity in graph.Therefore,through diving into knowledge graph,we can better comprehend user preference and provide better recommendation results for each user.The main work of this thesis is as follows:(1)The thesis introduces the development of recommendation system.We analyze different traditional recommendation algorithms and summarize their superiority and limitation.Thereafter we briefly introduce knowledge graph and shed light on knowledge graph based recommender systems.(2)We construct hybrid graph which integrates user-item interaction graph with item knowledge graph.Through diving into the constructed graph,users' collective behavior as well as latent relation between items can be discovered.(3)We apply graph neural network for representation learning on the extracted knowledge graph,through propagating information from neighbor nodes iteratively,precise representation can be learnt for each node in the graph.The information propagation process is relation aware in our model,which indicates that we not only make use of the node vector but also utilize the relation vector when information is propagating from one node to another.The final information vector the node to be updated receives is a combination of its neighbor vector and the relation vector between the node itself and its neighbor.The final vectors learnt for each user and item are utilized for downstream recommendation.(4)We conduct extensive experiments on Movie Lens and last.FM dataset and the results prove the efficiency of proposed method.(5)We develop a knowledge graph based recommender system,which implements the recommendation algorithm based on knowledge graph.The system can infer users' interest from his or her historical interactions and make recommendations for users.