Research On Graph-based Recommendation Algorithm

Due to tremendous profits and convenience for users and websites,recommender system has become an integral part of the Internet,and attracted many researchers to explore better recommendation algorithms.The key idea of recommendation algorithm is to characterize users and items,and collaborative filtering considers that there is tremendous information in the interaction history between users and items,from which we are model users' interest preferences and item characteristics.Although collaborative filtering has been achieved excellent performance,it still has some problems,such as inaccurate modeling of users and items,sparsity and cold start.The main content of this thesis is the research of recommendation algorithm based on graph structure,our research goal is to use graph structure and graph neural network to solve the above problems.This thesis mainly includes the following two parts:1)We propose a graph-based recommender algorithm with explicit feedback.Collaborative filtering algorithms usually represent the interaction history of users and items as a matrix,each update involves only one user and one item,and ignores the interaction history of users and items.Different from this representation,we construct a user-item interaction graph to describe the interaction history.In the interaction graph,the neighboring nodes of the node represent its interaction history.Therefore,we apply graph convolutional networks on interaction graphs to generate more expressive latent representations for users and items.Experiments show that our method achieve better performance than baseline method in rating prediction tasks.2)We propose a TOP-N recommendation algorithm based on knowledge graph.Auxiliary information has been introduced to collaborative filtering algorithms to alleviate the sparsity and cold start problems.As the most important knowledge representation method,knowledge graph can provide rich attribute information for items.We combine the leaning of knowledge graph with recommendation task,and on the basis of using the relation provided by the knowledge graph,fully exploits the relevance of nodes in the knowledge graph.thus take full advantage of knowledge graph in the recommendation process.Experiments demonstrate that our method is superior to the comparison methods in CTR prediction and recall tasks.