运用上下文增强和分层注意力机制的社交推荐  

Social Recommendation Based on Context Enhancement and Hierarchical Attention Mechanism

作　　者：孙克雷[1] 许峰[1] 周华平 SUN Kelei;XU Feng;ZHOU Huaping(School of Computer Science&Engineering,Anhui University of Science&Technology,Huainan 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：《厦门理工学院学报》2022年第5期35-43,共9页Journal of Xiamen University of Technology

基　　金：国家自然科学基金项目(61703005);安徽省重点研究与开发计划国际科技合作专项(20200,4b11020029)。

摘　　要：针对现有方法在建模用户、项目及其上下文特征上与社交网络耦合程度低,且没有充分挖掘社交信息中其他细粒度特征的问题,提出一种基于上下文增强和分层注意力机制的社交推荐模型(context enhancement and hierarchical attention mechanism social recommendation,CEHA-SR)。针对社交网络用户间关系所具有的图结构性质,该模型以图神经网络为框架,对社交信息、用户-项目-类别信息使用分层的注意力机制从不同层面的特征进行充分建模,并自适应得到不同特征之间的关系权重。在Ciao-28和Epinions-27两个真实数据集上的验证表明,该模型的均方根误差和平均绝对误差比经典的图神经网络社交推荐模型(GraphRec)分别降低了约3.63%、4.13%和4.33%、4.12%。To enhance the low coupling with social networks and insufficient mining of other fine-grained features in social information for the existing modeling of users,projects and their contextual features,a social recommendation model based on context enhancement and hierarchical attention mechanism(CEHA-SR)is proposed.Aiming at the graph structure nature of the relationship between social network users,the model takes graph neural network as the framework,and uses hierarchical attention mechanism for social information and user-item-category information to fully model features at different levels,and adaptively obtain the relationship weight between different features.The validation on two real data sets,Ciao-28 and Epinions-27,shows that the root mean square error and average absolute error of the model are reduced by 3.63%,4.13%,4.33%and 4.12%respectively compared with the classical graph neural network social recommendation model(GraphRec).

关 键 词：社交推荐模型 注意力机制 社交融合 图神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]