基于梯度感知图增强的图对比学习推荐算法

打开文本图片集

关键词：推荐系统；图对比学习；图增强；梯度感知；难负样本

中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2026）01-016-0136-09

doi：10.19734/j. issn.1001-3695.2026.05.0172

Graph contrastive learning recommendation algorithm based on gradient-aware graph augmentation

Hu Jichao a，b† ，Hu Xiyao°，Li Huanzhea （a.ColegeoffoaionEgneing，tellgentSoeokEneiRseenteofbernc，Hbe Geit Shijiazhuang 050031，China）

Abstract：Toaddress the problemthatrandomaugmentations ingraph-basedcontrastiverecommendationmodels tend tocause semantic informationlossand intensifypopularitybiasedly，thispaperproposedanovelcontrastivelearningrecommendation algorithmbasedonGAA.Fistly，itintroducedgradient-awareaugmentationstrategy，whichusedgradientinformationtoguide theconstructionofcontrastive viewsthat beter preservedsemantic structures.Secondly，itemployed a hard negative mining methodbasedonpositive similarity toenhancethe model’sabilitytodistinguish importantsemanticsandreducepopularity bias.Finall，itincorporatedatemperature-adaptivemoduletoensuretrainingstability.Itconductedexperimentsonthree publicdatasets，including Yelp2O18，Amazon-book，and Alibaba-iFashion，with evaluation metricsRecall@ K and NDCG @ （2 （204号 K ，On the Alibaba-iFashion dataset，GAA achieved 7.94% and 5.82% improvements in Recall@ K and NDCG@ K ，respectively，comparedwiththebestbaseline.TheresultsverifytheeffectivenesofGAAinallviatingsemanticinformationlosand popularity bias.

Keywords：recommender system；graph contrastive learning；graph augmentations；gradient awarenes；hard negative samples

0 引言

随着网络信息技术的发展，信息量激增，用户难以筛选所需内容，推荐系统由此产生，其用于缓解信息过载问题[1]。（剩余22056字）