基于对齐性和均匀性约束的图神经网络会话推荐方法  

Graph neural network toward representation alignment and uniformity for session-based recommendation

作　　者：唐韬韬 楚飞 汪炅 贾彩燕[1,2] TANG Taotao;CHU Fei;WANG Jiong;JIA Caiyan(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044,China;Beijing Key Lab of Traffic Data Analysis and Mining,Beijing 100044,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]交通数据分析与挖掘北京市重点实验室,北京100044

出　　处：《应用科技》2024年第2期90-98,共9页Applied Science and Technology

基　　金：国家重点研发计划项目(2018AAA0100302).

摘　　要：会话推荐(session-based recommendation,SBR)旨在匿名状态下通过用户的短期历史行为序列来预测下一个待点击的项目。为解决现有基于图神经网络(graph neural networks,GNNs)的会话推荐方法忽略会话中不同位置相同项目之间差异的问题,在图卷积获得项目表示后,进一步考虑该项目与相邻项目之间的相关性,生成邻域相关的项目表示。此外,鉴于对齐性和均匀性在对比学习中的起到的重要作用,还提出了一种适用于会话推荐的对齐性和均匀性损失方法,以约束生成的会话表示和项目表示。在3个公开数据集上的实验表明,文中提出的模型TAU-GNN的推荐性能优于对比的主流会话推荐模型。Session-based recommendation(SBR)aims to predict the next click item of an anonymous user based on a short interaction sequence.In order to solve the problem that existing SBR methods based on graph neural networks(GNNs)ignore the differences between the same items at different locations in a session,after obtaining an item embedding by a GNN,we further consider the correlation of its neighbor items in a session to generate the item neighborhood correlation representation.Considering the importance of alignment and uniformity in contrastive learning,we propose an alignment and uniformity loss method for session recommendation to constrain the generated session representation and item representation.Experiments on three public benchmark datasets show that our newly proposed model TAU-GNN is better than that of the mainstream models for SBR.

关 键 词：会话推荐 图神经网络 对齐性 均匀性 对比学习 交叉熵损失 匿名会话 邻域信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]