基于分离式表征的知识图谱推荐算法  

Knowledge Graph Recommendation Algorithm Based on Disentangled Representation

作　　者：陈清泓 林广艳[1] 柳啸峰 谭火彬[1] CHEN Qinghong;LIN Guangyan;LIU Xiaofeng;TAN Huobin(School of Software,Beihang University,Beijing 100191,China)

机构地区：[1]北京航空航天大学软件学院,北京100191

出　　处：《武汉大学学报（理学版）》2021年第6期517-524,共8页Journal of Wuhan University:Natural Science Edition

基　　金：国家重点研发计划(2018YFB1402600)。

摘　　要：基于知识图谱的推荐算法可以挖掘用户的潜在兴趣,有效解决推荐系统中的冷启动和数据稀疏性问题,然而现有基于知识图谱的推荐算法建模层面单一,很难挖掘到用户的深层兴趣。为此,提出一种基于分离式表征的知识图谱推荐算法。该算法首先利用分离式表征方法,将用户和物品的混合表征解耦成多个层面的分离式表征;然后采用图神经网络方法,利用用户-物品交互二分图和知识图谱中的邻域信息扩充用户和物品的分离式表征;同时,在分离式表征聚合邻域信息时,采用注意力机制和门控单元区分不同信息的重要性,自适应捕捉用户兴趣点。该算法细粒度刻画用户和物品表征,深度挖掘用户兴趣和物品特征,在三个公开数据集上与基准算法进行了对比实验,实验结果表明本文提出的算法在AUC、F1等指标上有明显提高。Knowledge graph-based algorithms can explore users’ potential interest, alleviating the cold start and data sparsity issues in recommender systems. However, existing relevant algorithms are difficult to capture users’ deep interest due to the coarse-grained modeling. In this study, we propose a knowledge graph recommendation algorithm based on disentangled representation. Firstly, the coupled representation of users and items is decoupled into several disentangled representations. Then,the graph neural network is introduced to augment the disentangled representation by aggregating neighborhood information in user-item bipartite graph and knowledge graph. During aggregation, the attention mechanism and the gate unit are adopted to discriminate the importance of different neighbors. We conduct comparative experiments with baselines on three public benchmark datasets. The improvement under several metrics like AUC and F1 demonstrates the superiority of our proposed algorithm in real scenarios.

关 键 词：知识图谱 推荐系统 分离式表征 图神经网络 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]