Research On Bundle Recommendation Algorithm Based On Graph Convolutional Networks And Contrastive Learning

In real life,e-commerce platforms often use bundled sales as a marketing strategy to increase sales,recommending combinations of multiple items to users.Traditional recommendation models typically focus on recommending single items to users,which is difficult to meet the above needs.Therefore,research on bundle recommendations has received widespread attention from academia and industry.Unlike traditional recommendation models,bundle recommendations require recommending bundles containing multiple items to users,which often face more serious data sparsity problems and involve multiple complex relationships among users,items,and bundles.This paper aims to delve into these issues and provide users with more accurate bundle recommendations.Firstly,to address the data sparsity problem in bundle recommendation,this paper proposes a graph convolutional bundle recommendation model based on meta-path,which incorporates user’s additional interests in items into bundle recommendation and establishes a user-item-bundle tripartite graph.Meanwhile,to reduce the interference of user-item interaction paradigms on learning high-order similarity information between users and bundles,this paper establishes user and bundle adjacency graph based on meta-paths in the tripartite graph.By applying graph convolutional networks on the two types of graphs,different high-order features of users and bundles are captured to improve the accuracy of bundle recommendation.Secondly,to address the complex relationships among multiple entities in bundle recommendation and the lack of clear relationships between users and bundles in the graph convolutional bundle recommendation model based on meta-path,this paper proposes a bundle recommendation model based on relational graph with contrastive learning.This model analyzes two different interaction modes between users and bundles: direct and indirect.For direct interactions,a direct relationship graph is established using the interaction history between users and bundles.For indirect interactions,considering the interaction history between users and items as well as the inclusion history between bundles and items,an indirect relationship graph is established using items as bridges.Meanwhile,this paper uses contrastive learning to model the cooperative associations between the two different relationship graphs and further improves the accuracy of bundle recommendations by adding uniform distribution noise data augmentation methods.Finally,the two models are validated on real-world datasets.Comparative experiments and ablation experiments are designed to prove the effectiveness and rationality of the proposed models.