基于图网络嵌入及BPR的推荐算法  

Recommendation algorithm based on graph embedding and BPR

作　　者：吴彪[1] 温雯[1] 郝志峰[1,2] 蔡瑞初 WU Biao;WEN Wen;HAO Zhi-feng;CAI Rui-chu(School of Computers,Guangdong University of Technology,Guangzhou 510000,China;School of Mathematics and Big Data,Foshan University,Foshan 528225,China)

机构地区：[1]广东工业大学计算机学院,广东广州510000 [2]佛山科学技术大学数学与大数据学院,广东佛山528225

出　　处：《计算机工程与设计》2021年第6期1638-1644,共7页Computer Engineering and Design

基　　金：NSFC-广东联合基金项目(U1501254);国家自然科学基金项目(61976052、61876043)。

摘　　要：通过构建一个能够体现不同用户-项目交互信息和项目属性信息的图网络,将推荐问题转化为异构图网络的节点嵌入问题并提出一种用于生成用户-项目表示的联合学习算法。在该联合学习算法中,采用带权的个性化排序算法体现用户对于不同项目的喜好程度的差异,融合项目信息拓展用户的行为模式,更好地找到用户的喜好。实验结果表明,该算法在多种不同的推荐评价指标上比基准算法具有明显的优势。By constructing a graph network which is able to integrate different user-item interactions and item information,the recommendation problem was transformed into the node embedding of heterogeneous graph network.A joint learning algorithm was developed for producing user-item representations.In this joint learning algorithm,personalized ranking algorithm with weights was used to reflect the difference of users’preferences for different items.It was worth noting that by expanding the user behavior pattern through integration of item information,it was able to better predict user preferences.Experimental results show that the proposed algorithm has significant advantages over the state-of-the-art baselines.

关 键 词：推荐算法 表示学习 图网络嵌入 隐式反馈 个性化排序 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]