Research On Graph Convolutional Collaborative Filtering Algorithm Under Multiple Relations

With the popularity of Internet applications,the information in the network has exploded and the problem of information overload has become increasingly serious.How to design efficient recommendation algorithms to filter out items of interest for users has been a hot research topic.Graph convolutional Network(GCN)have achieved great success in recent years due to their ability to capture higher-order relationships between entities,and have become the state-of-the-art of collaborative filtering.Nevertheless,existing GCN-based collaborative filtering algorithms still have the following problems:(1)only mining the history interaction of users and items,which only reflects users’ preferences from one aspect;(2)only acting on the data of bipartite graph structure without modeling the relationship between the same type of entities in explicit manner.This paper proposes corresponding improvements to address the above problems,and the main work is as follows.(1)We propose a Light GCN based Aspect-level Collaborative Filtering(LGCACF)algorithm,which is a multi-module training model.First,we construct aspect-level interaction graphs based on user-item interaction history and item knowledge information,and adopt GCN to act on these interaction graphs separately to learn the embedding vectors of different aspects of users and items;then,we adopt two layer Full Connection to fuse the embeddings of different aspects,which can endow the interaction model with nonlinear transformations and effectively extract nonlinear collaborative signals of aspect-level interactions of users and items.(2)We propose a User and Item Relation Enhanced Graph Convolutional Collaborative Filtering(UIGCCF)algorithm,which construct user relationship graphs and item relationship graphs using rating similarity or tag information and then propagates and aggregates the embeddings in the heterogeneous relationship graphs.The algorithm expands the original bipartite graph topology by exploring the association of the same type of entities to capture more complex collaborative signals.(3)Numerous experiments are conducted on public datasets,and the experimental results show that compared with other start-of-the-art algorithms,the two models proposed in this paper achieve significant improvement under the three metrics of precision,recall,and NDCG,and alleviate the data sparsity problem to a certain extent.