基于大语言模型与图神经网络的会话推荐增强框架

打开文本图片集

关键词：会话推荐；大语言模型；图神经网络；个性化推荐；语义增强；迭代去噪中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2026）01-004-0035-08doi：10.19734/j.issn.1001-3695.2025.06.0201

Session-based recommendation enhancement framework based on large language models and graph neural network

YuEnhai，WenYan†，Chen Yuao （ColegeofComputerScienceandEngineering，handong UniversityofScienceandTechnologyQingdaoShandong2669China）

Abstract：WiththewidespreadapplicationofSBR，itisthecriticalchalengesofutilizingsemanticinformation，modeling cross-sessionuser interests，andsuppresingdatanoise toimproverecommendation perfor-mance.This paperproposedLGBR，a novel framework that integrateslargelanguage models （LLMs）and graph neural networks（GNNs）toachieve semantic enhancementandpersonalizedrecommendation.Specificall，itgeneratedsupplementarytextembedngsforitemsandcrosssessionuser interest embeddngsusingLLMsand fine-tunedlanguage models，fusedthese withIDembeddingsthroughasoft atentionmechanismtocreatesemanticallrichrepresentations，incorporateduserinterestembeddingswithalignmentlossfor personalizedrecommendations，andapplied two-stageweightlearning tofilternoisyitemsandoptimizesessionrepresentations. Experiments demonstrate that LGSBR achieves P@20 of 21.38% and MRR@20 of 6. 76% on the Beauty dataset，improving over the SR-GNN baseline by 23.3% and 50.56% ，respectively，and P@20 of 25. 86% and MRR@20 of 7.58% on the Movie Len- Ω1M dataset，with gains of 12.63% and 10.98% .The study confirms LGSBR’s generality and effectiveness across multiple GNN models.

Key words：sesion-based recommendation（SBR）；large language model；graph neural network；personalized recommendations；semantic enhancement； iterative denoising

0 引言

随着互联网的快速发展，推荐系统[12]在电子商务和社交媒体等领域中的作用越来越重要。（剩余19547字）