基于分级注意力网络和多层对比学习的社交推荐  

Multi-level Attention Networks and Hierarchical Contrastive Learning for Social Recommendation

作　　者：张丽杰 王绍卿[1] 张尧 孙福振[1] ZHANG Lijie;WANG Shaoqing;ZHANG Yao;SUN Fuzhen(School of Computer Science and Technology,Shandong University of Technology,Zibo Shandong 255000,China)

机构地区：[1]山东理工大学计算机科学与技术学院,山东淄博255000

出　　处：《广西师范大学学报(自然科学版)》2025年第2期133-148,共16页Journal of Guangxi Normal University:Natural Science Edition

基　　金：山东省自然科学基金(ZR2020MF147,ZR2021MF017)。

摘　　要：将社交关系融入推荐系统中,能有效提高推荐质量。然而现实世界中用户的交互数据是稀疏和复杂的,如何更好地利用社交信息是关键问题。现有社交推荐模型没有充分探索高阶好友的影响,而且忽略了用户间的关系强度和不同种类的关系对用户的影响,导致推荐性能不佳。为了解决上述问题,本文提出一个基于分级注意力网络和层次化对比学习的社交推荐模型。具体来说,首先,依据用户间不同关系构建用户级超图,扩大节点聚合的感知范围,加深模型深度。然后,设计多级注意力网络更好地捕捉用户交互数据之间的关系和重要性,其中,视图级自注意力机制捕获好友对用户的影响以及项目间的关联程度,通道级注意力自适应地调整不同种类的关系对用户的影响。同时,引入层次化对比学习对数据进行增强,包括视图间和跨视图的第一层对比学习和针对高阶关系的第二层对比学习,多维度捕获数据的细微差距和高层次的抽象特征。最后,将所提出的模型在4个公开基准数据集上进行评估,结果表明本文模型Precision、Recall、NDCG较其他最优基线模型分别提升7.61%、11.05%、10.69%,验证了本文模型的有效性。Incorporating social relationships into recommender systems can effectively improve recommendation quality.However,real-life interactions among users are sparse and complex.The effective utilization of social information is akeyissue.The impact of high-order friends is not fully explored by existing social recommendation models,and the strength of relationships between users and the impact of different kinds of relationships on users are ignored,leading to sub-optimal recommendation performance.To address these issues,Multi-level attention networks and Hierarchical Contrastive Learning for social recommendation(MHCL)is proposed.Specifically,the user-level hypergraph is first constructed based on different relationships between users to expand the perceptual scope of node aggregation and deepen the depth of the model.Then,a multilevel attention network is designed to better capture the relationships and importance between user interaction data,where the influence of friends on the user and the degree of inter-item relatedness are captured by the view-level self-attention mechanism,and the influence of different kinds of relationships on the user is adaptively adjusted by the channel-level attention.Meanwhile,hierarchical contrast learning is introduced to augment the data,including the first level of contrast learning between and across views and the second level of contrast learning for high-order relationships,to capture the subtle gaps and high-level abstract features of the data in multiple dimensions.Finally,the proposed model is evaluated on four publicly available benchmark datasets,and the effectiveness and reasonableness of MHCL are validated by the evaluation results.Social denoising will be the focus of future research to improve recommendation systems based on hypergraph neural networks.

关 键 词：社交推荐 注意力网络 超图学习 对比学习 推荐系统 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]