基于增强图神经网络的特征融合会话推荐方法  

Feature fusion session recommendation method based on enhanced graph neural network

作　　者：袁凤源 梅红岩 温民伟 白杨 YUAN Feng-yuan;MEI Hong-yan;WEN Min-wei;BAI Yang(School of Electronics and Information Engineering,Liaoning University of Technology,Jinzhou 121001,China;School of Software,Liaoning University of Technology,Jinzhou 121001,China)

机构地区：[1]辽宁工业大学电子与信息工程学院,辽宁锦州121001 [2]辽宁工业大学软件学院,辽宁锦州121001

出　　处：《计算机工程与设计》2025年第2期546-553,共8页Computer Engineering and Design

基　　金：国家自然科学基金面上基金项目(62273170);辽宁省教育厅面上基金项目(JYTMS20230869);辽宁省教育厅科研基金项目(JZL202015404、LJKZ0625)。

摘　　要：针对现有会话推荐方法过于依赖局部特征的交互提取,忽略全局特征的关联性,导致会话上下文信息丢失的问题,提出一种基于增强图神经网络的特征融合会话推荐方法(FFSR-EGNN)。引入卷积单元到图神经网络中,更有效捕获不同层次的局部特征,采用潜在语义分析方法构建会话-项目矩阵提取会话的全局特征,利用注意力机制自适应地学习局部和全局特征,使模型能够根据特征节点的关联性调整、增强节点间的重要信息传递。通过线性部件将局部和全局特征进行融合,生成最终的语义表示和预测概率用于推荐任务。在Diginetica和Yoochoose数据集上实验,其结果表明,所提方法推荐性能优于现有主流推荐方法。To address the issue of context information loss in existing conversation recommendation methods that rely heavily on local features and overlook the correlation of global features,a session recommendation method based on enhanced graph neural network(FFSR-EGNN)was proposed.By introducing convolution units into the graph neural network,different levels of local features were more effectively captured.The latent semantic analysis method was used to construct a session-item matrix to extract the global features of the session,and the attention mechanism was utilized to adaptively learn local and global features,enabling the model to adjust and enhance the important information transmission between feature nodes according to their correlation.The local and global features were fused through linear components to generate the final semantic representation and prediction probability for the recommendation task.Experimental results on the Diginetica and Yoochoose datasets show that the proposed method outperforms existing mainstream recommendation methods.

关 键 词：会话 推荐系统 深度学习 机器学习 图神经网络 特征融合 会话推荐 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]