基于完整超图神经网络的捆绑推荐模型

打开文本图片集

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）07-011-2003-08

doi：10.19734/j.issn.1001-3695.2024.11.0494

Abstract：Bundling recommendation enhances user experienceand boosts merchant sales performance byofering predefined setsof productcombinations.It also playsa significantrole in various serviceecosystems suchas video-on-demandandmusic playlistgeneration.Existingbundlingrecommendationmethodsoftenrelyonsharedmodel parametersormulti-task learning schemes，neglectingthedeep-levelconnectionsamongusers，items，andbundles，hichleadstoinformationlossandipacts the performanceofrecommendationsystems.Toadesstheseissues，tispaper proposedacompletehypergraphneuralnetwork （CHNN）.Firstly，theframeworkconstructedacompletehypergraphtoexpressthetemaryelationshipsamong users，items，nd bundles.Theseternaryrelationshipsnotonlyincludedtheinterconnectionsamongusers，items，andbundled，butalsoecompasedthe interalconnectionswithinusersandbundles，effectivelydescribingtherelationshipbetweenproductbundlesand userpreferences.Scondly，temodelconsistsofaninitializationlayer，atripleconvolutionlayer，andapredictionlayer.The initializationlayergeneratedembeddingvectorsforeachuser，item，andbundle.Thetripleconvolutionlayerextractediforationfromthecompletehypergraphandleveragedtheuser-bundle graphand item-bundlegraph toenhancetherepresentations of users，items，andbundles.The predictionlayer providedrecommendationsbasedonthefinalembedding vectors.Through multi-layerrichconvolutionoperations，themodel fullyexploredtheassciationscontained inthecomplete hypergraphto achieve moreaccuraterecommendations.Experimentson tworeal-worlddatasets，NetEaseandYouShu，demonstratethat CHNN achieves an average improvement of 2.4% in recall and 2.75% in NDCG，outperforming existing baseline models and showcasing its effectiveness in the field of bundling recommendation.

Key words：graph neural network；bundle recommendation；hyper graph；graph convolutional network

0 引言

随着营销策略的快速发展，商家已经考虑向用户推荐一组商品[1-5]。（剩余19197字）