项目邻居信息对比增强的推荐方法  

Recommendation Method for Contrastive Enhancement of Item NeighborhoodInformation

作　　者：周北京 王海荣 马赫 张丽丝 ZHOU Beijing;WANG Hairong;MA He;ZHANG Lisi(College of Computer Science and Engineering,North Minzu University,Yinchuan 750021,China;The Key Laboratory of Images&Graphics Intelligent Processing of State Ethnic Affairs Commission,North Minzu University,Yinchuan,750021,China)

机构地区：[1]北方民族大学计算科学与工程学院,宁夏银川750021 [2]北方民族大学图像图形智能处理国家民委重点实验室,宁夏银川750021

出　　处：《山西大学学报（自然科学版）》2024年第2期269-278,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：宁夏自然科学基金(2023AAC03316);北方民族大学研究生创新项目(YCX23146;YCX23159)。

摘　　要：针对基于知识图谱推荐的方法中存在的用户-项目交互监督信号弱,以及知识图谱中包含噪声信息的问题。本文提出了一种项目邻居信息对比增强的推荐方法-RMCEIN(A Recommendation Method for Contrastive Enhancement of Item Neighbor Information)。RMCEIN通过异构传播和知识感知注意力函数获得用户和项目的多阶邻居嵌入,用于丰富用户和项目的特征;在项目邻居嵌入过程中采用添加均匀分布的弱噪声的方式,构建项目邻居增强视图,可以有效地减少视图构建的时间开销。此外,通过两个项目邻居视图间的对比学习,调用对比损失函数促使项目视图信息的均匀性,调节项目的邻居结构,达到减少知识图谱中知识噪声的目的,同时引入多任务学习来缓解监督信号弱的问题。为了验证方法的有效性,在三个公共数据集上进行实验,将实验结果与10种基于知识图谱的推荐方法进行了对比,本文方法的AUC(Area Under Curve)平均提升了2.32%,F1值平均提升了2.26%。Aiming at the weak user-item interaction supervision signal in the knowledge graph-based recommendation method and the problem that the knowledge graph contains noise information,in this paper,we propose a recommendation method for contrastive enhancement of item neighbor information(RMCEIN).The RMCEIN obtains the multi-order neighbor embedding of users and items through heterogeneous propagation and knowledge-aware attention function,which is used to enrich the characteristics of users and items;in the process of item neighbor embedding,it adopts the method of adding uniformly distributed weak noise to construct item neighbor enhancement view,which can effectively reduce the time overhead of view construction.In addition,through contrastive learning between two item neighbor views,the contrastive loss function is called to promote the uniformity of item view information,adjust the neighbor structure of items,achieve the purpose of reducing knowledge noise in the knowledge graph,and at the same time introduce multi-task learning to alleviate the supervision signal weak problem.In order to verify the effectiveness of the method,experiments were carried out on the MovieLens-1M,Book-Crossing and Last-FM datasets,and the experimental results were compared with 10 methods such as RippleNet (Propagating User Preferences on the Knewledge Graph for Recommender Systems),CKAN, KGIC, etc. The AUC (Area Under Curve) of the method in this paper increased by 2.32% on average. F1 value increasedby 2.26 on average.

关 键 词：知识降噪 多任务学习 知识图谱 推荐方法 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]