融合图神经网络和稀疏自注意力的会话推荐分析  

Session-based Recommendation Analysis of Fusing Graph Neural Networks and Sparse Self-Attention

作　　者：胡胜利[1] 程春 HU Shengli;CHENG Chun(School of Computer Science and Engineering,Anhui University of Science&Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《兰州工业学院学报》2023年第6期13-18,共6页Journal of Lanzhou Institute of Technology

基　　金：安徽理工大学研究生创新基金项目(2022CX2121)。

摘　　要：针对现有会话推荐采用单一模型无法兼顾全局和局部信息,从而影响推荐性能的问题,提出融合图神经网络和稀疏自注意力的会话推荐模型(SSA-GNN)。模型采用稀疏自注意力构建全局隐向量,以解决无关项的干扰和图神经网络难以表示长距离依赖的问题;采用目标注意图神经网络构建局部隐向量,更深层次的捕获项目间的复杂依赖。最后在预测层将全局和局部隐向量线性连接,有效兼顾了全局和局部信息。模型在Yoochoose1/64数据集上的试验结果比基线模型GC-SAN在评价指标P@20上提高了1.25%,MRR@20上提高了4.59%。Aiming at the problem that the existing session recommendation model cannot take into account the global and local information,which affects the recommendation performance,a session recommendation model(SSA-GNN)combining graph neural network and sparse self-attention is proposed.The model uses sparse self-attention to construct a global hidden vector to solve the interference of irrelevant items and the difficulty of representing long-distance dependencies in graph neural networks;Using a target attention graph neural network to construct local hidden vectors and capture complex dependencies between items at a deeper level.Finally,the global and local hidden vectors are linearly connected in the prediction layer,effectively balancing global and local information.Compared with baseline model GC-SAN,the experimental results on Yoochoose1/64 dataset have increased by 1.25%on P@20 and 4.59%on MRR@20.

关 键 词：会话推荐 图神经网络 稀疏自注意力 目标注意 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]