Bundle Recommendation Model Based On Hypergraph Attention And Contrastive Learning

In modern business,bundle sales are one of the most important marketing strategies for promoting products by merchants.Bundle recommendation is a method of recommending a bundle,which is a collection of items,to users.Many current works use graph neural networks to model interaction information between entities,and propagate embedding features by modeling user intent and bundle composition.However,this approach has some limitations.Firstly,graph network modeling is aimed at pairwise relationships,which may lead to inadequate encoding of higherorder association information.Secondly,the interaction between users and bundles in bundle recommendation is more sparse than item recommendation,which requires enhancing the representation learned by the model.Thirdly,in bundle recommendation,two aspects are important: how to represent user interests and how to reflect the membership relationship between bundles and items.To address these issues,this paper proposes a bundle recommendation method based on hypergraph attention and contrastive learning.The following work is done in this paper:The research background of bundle recommendation and some current research methods are introduced.For non-pairwise relationships in bundle recommendation,a hypergraph neural network is used in this paper,and an attention mechanism is added on the basis of the hypergraph convolutional network to represent the probabilistic aggregation in the two-stage propagation process.A bundle recommendation model based on hypergraph attention and contrastive learning is proposed.The user interest hypergraph is modeled with the user as the center,using hyperedges to build user interest relationships.The bundle hypergraph is modeled by considering the internal differences and external similarities between bundles,using hyperedges to build the membership relationship between bundles and items,and the internal differences between bundles.Shared elements are used to express the similarity between bundles.The contrastive learning method is used to reduce the impact of sparsity on the representations obtained by cross-view propagation.A bundle recommendation model based on multi-channel hypergraph and contrastive learning is proposed.After analyzing the data of bundle recommendation,three tasks are proposed: user-bundle task,user-item task,and bundle-item task.Three hypergraphs are used to establish channels for users’ interests in bundle packages,channels for users’ interests in items,and channels for bundle packages’ selection of items.The first two channels express two aspects of users’ interests,and the third channel contains collaborative filtering signals where items with similar features are gathered into common bundle packages.Contrastive learning is used to enhance the representation of embeddings obtained through cross-view aggregation.This paper adopts a joint training strategy of multi-task learning,and experiments on hypergraph representation learning,cross-view collaboration,and self-supervised contrastive learning methods of the model,and verifies the effectiveness of hypergraph representation learning,cross-view collaboration,and self-supervised contrastive learning methods.