融合社交影响扩散的长尾物品推荐模型

打开文本图片集

中图分类号：TP301 文献标志码：A 文章编号：0253-2395（2025）04-0741-11

Abstract： Inrecommendationsystems，thelong-taildistributionofuserratingsandinteractionfrequenciesposeschallnges forextracting thefeaturesoflong-tail items.Existing methods eitheroverlyfocusontailitems whileneglectingteirconnections with headitemsordisregard theinfluenceofsocialnetworksonuserpreferences，therebyimpactingrecommendationperformance.Toaddrestheseissues，thispaperproposesanovellong-tailrecommendationmodelcaledLoSidi （Long-tailRecommendationMethod IncorporatingSocialIfluenceDifusio）.Firstlyforachusethmodelagegatessamplesofocialeigorsatvariola andintegratesthesewiththepopularitemstheuserhasinteractedwithtogenrateuserinterestembeddings.Secondlythepotential featuresoflong-tail temsareminedbycalculatingthesimilaritybetweenlong-tailitemsandthehaditems theuserhasinteracted with.FinalytheLoSidimodelstablisheslinksbetwenusersandlong-tailitems，predictingscorsandeneratingrecoedationsfortheseitems.Experimentalresultsonwidely-useddatasetsdemonstratethattheproposed modelsignificantlyimproves the novelty and diversity of users'recommendation lists.

Key words： recommender system; social recommendation; long tail problem; graph neural network

0 引言

推荐领域实证研究表明，用户行为数据（包括评分记录和交互频率）普遍遵循长尾分布[1-2]，揭示出系统内存在海量具有潜在价值的非流行物品（即长尾物品）亟待有效识别[3]然而，基于协同过滤的经典推荐模型虽然在头部物品推荐中展现出良好的预测性能，但对于长尾物品的精准推荐却面临推荐覆盖率与准确率的双重困境。（剩余15333字）