Rule-Guided Graph Neural Networks For Recommender System

The classic collaborative filtering algorithm has achieved excellent results in rec-ommender system.However,because the collaborative filtering algorithm is only fo-cused on exploring the history of user-item interactions,it is difficult to perform well in a recommender system where user-item interactions are sparse,i.e.,the collaborative filtering algorithm cannot properly handle the cold start problem.In order to alleviate the cold start problem that the collaborative filtering algorithm is unable to deal with,more and more recommendation algorithms begin to use external knowledge,such as the user's own attributes,to provide information support for the algorithm.As a high-quality structured data,the knowledge graph is being used by more and more recom-mender systems.Existing recommendation algorithms that utilize knowledge graphs can be roughly divided into three categories: embedded-based,path-based(rule-based),and aggregation-based.These three types of algorithms have achieved certain results in practice,but also have some defects.Recommendation algorithms based on em-bedding usually use mature knowledge graph embedding models to obtain knowledge graph embedding.They only consider the direct connections between users and enti-ties,and ignore the long-range semantic dependence between users and entities.The recommendation algorithm based on the paths(rules)can use the paths to capture long-range semantic dependence,but the linear modeling method guided by rules cannot fully consider the diversified connections between different entities.The aggregation-based method fully considers the diverse connections between entities by designing a unique aggregation mechanism or using graph neural networks,but still has problems in capturing explicit long-range semantic dependence.In order to capture the long-range semantic dependence between users and en-tities while fully considering the diverse connections between different entities,this paper combines rules and graph neural networks to design recommendation algorithm RGRec(Rule-Guided Graph Neural Networks for Recommender System).In the rules section,in order to obtain high-quality rules,this paper designs automatic rule mining and fast rule fitlering methods,and uses rules to capture long-range semantic depen-dence between users and entities.Regarding the graph neural network,for the problem of information loss and noise generating caused by selecting points randomly,this ar-ticle uses rules to guide the point selection,and establishes clear semantic connections between the central nodes and the remote nodes,and obtains the multi-dimensional fea-tures guided by the rules,and through the weighted integration of the multi-dimensional features,the information around the central node is reasonably gathered to fully con-sider the connections between different entities and prevent the entities from being treated in isolation.This paper designs comparative experiments.In the two scenarios of click-through rate prediction and top-K recommendation,three datasets Last.FM,Movie Lens-1M and Dianping-Food are used to compare RGRec with multiple repre-sentative methods.The comparison of the different algorithms proves the effectiveness of RGRec,and also proves that it is very successful to combine rules and graph neural networks.