融合全局信息的多图神经网络会话推荐  

Global Information Multi-graph Neural Network for Session-based Recommendation

作　　者：黄涛 徐贤[1] HUANG Tao;XU Xian(Department of Computer Science and Engineering,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学计算机科学与工程系,上海200237

出　　处：《小型微型计算机系统》2024年第4期769-776,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61872142,62072299)资助;上海市高可信计算重点实验室开放课题项目(OP202205)资助.

摘　　要：基于会话的推荐旨在根据当前会话预测下一个最可能交互的物品.由于单个会话点击序列较短,仅使用会话本身的信息很难提供准确的推荐.因此,综合考虑其它会话之间的交互信息已成为一种趋势,为了提高推荐性能,本文提出一种融合全局信息的多图神经网络会话推荐模型(GIMGNN)来增强会话推荐的效果.该模型首先通过超图卷积神经网络和门控图神经网络从全局会话超图和局部会话图中学习两个级别的物品表示,然后通过注意力机制将反向位置信息融合到两种表示中,最后利用融合后的表示完成预测.在两个真实数据集Yoochoose和Diginetica上进行了一系列实验,实验结果表明,对比性能最优的基准模型,GIMGNN模型在Yoochoose上P@20和MRR@20至少提升了2.42%和4.01%,在Diginetica上P@20和MRR@20至少提升了6.56%和9.11%,验证了模型的有效性.Session-based recommendation aims to predict the next most likely item to be interacted based on the current session.Given the short duration of a single session and the short sequence of clicks,it is difficult to provide accurate recommendations using only information about the session itself.It has become a trend to comprehensively consider the interaction information between other session sequences.In order to improve the effect of session recommendation,Global information multi-graph neural network for session-based recommendation(GIMGNN)is proposed to enhance the effect of session recommendation.Firstly,the model obtains the global session representation and the local session representation of the item by Hypergraph convolutional neural network and Gate graph neural network respectively,and then integrates the reverse position information into the two representations by attention mechanism,and finally uses the fused representation to complete the prediction.A series of experiments have been carried out on two real data sets,Yoochoose and Diginetica.The experimental results show that compared with the benchmark model with the best performance,GIMGNN model is on Yoochoose P@20 and MRR@20 At least 2.42%and 4.01%were improved on Diginetica P@20 and MRR@20 At least 6.56%and 9.11%were improved,which verify the validity of the model.

关 键 词：会话推荐 超图卷积神经网络 门控图神经网络 注意力机制 位置信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]