结合注意力机制的图神经网络会话推荐  

AAttention Mechanism Graph Neural Network for Session-base Recommendation

作　　者：郑小丽 王巍[1,2,3] 张闯 杜雨晅 ZHENG Xiao-li;WANG Wei;ZHANG Chuang;DU Yu-xuan(College of information&Electrical Engineering,Hebei University of Engineering,Handan 056038,China;Hebei Key Laboratory of Security&Protection Information Sensing&Processing(Hebei University of Engineering),Handan 056038,China;College of internet of Things Engineering,Jiangnan University,Wuxi 214122,China)

机构地区：[1]河北工程大学信息与电气工程学院,河北邯郸056038 [2]河北省安防信息感知与处理重点实验室(河北工程大学),河北邯郸056038 [3]江南大学物联网工程学院,江苏无锡214122

出　　处：《电脑与信息技术》2023年第6期5-9,共5页Computer and Information Technology

摘　　要：会话推荐旨在根据用户历史行为序列预测将要与之交互的下一项。如何深入分析会话序列内部复杂的依赖关系,精确提取用户潜在偏好是目前会话推荐模型设计面临的巨大挑战。针对此,提出一种结合注意力机制的图神经网络会话推荐模型(AMSR-GNN)。该模型首先将全部会话数据构建为图,通过图神经网络获取图上节点局部嵌入表示;其次,利用含有噪声滤除器的注意力机制显示地过滤掉不重要的节点表征,得到去噪增强的全局嵌入表示,最后,由预测层考虑物品的局部嵌入表示和全局嵌入表示,为用户生成个性化推荐。Session recommendations were designed to predict the next item they will interact with based on a sequence of historical user behavior.How to deeply analyze the complex dependencies within the conversation sequence and accurately extract the potential preferences of users is a huge challenge in the design of the current conversation recommendation model.In view of this,a graph neural network session recommendation model(AMSR-GNN)combined with attention mechanism was proposed.Firstly,all session data was constructed as a graph,and the local embedding representation of nodes on the graph was obtained through the graph neural network.Secondly,the attention mechanism containing the noise filter was used to display and filter out the unimportant node representations to obtain the denoising-enhanced global embedding representation,and finally,the local embedding representation and global embedding representation of the item are considered by the prediction layer to generate personalized recommendations for users.

关 键 词：会话推荐 图神经网络 注意力机制 噪声滤除器 去噪增强 用户偏好 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]