摘要
利用知识图谱和深度学习进行推荐的方法得到了广泛的研究和应用,但是大多数推荐模型对物品表示建模不够完整,也未能全面捕捉和充分利用用户及物品的高阶交互信息。针对用户特征和物品特征高阶表示的提取问题,对用户与物品间的交互信息和知识图谱的关联信息进行联合提取,提出一种知识图谱交叉涟漪网络(KGCRN)。利用涟漪网络传播对用户偏好和物品特征进行建模,同时丰富两者的表示,提升推荐的性能。此外,设计一种改进的交叉压缩单元处理涟漪网络的输出,利用涟漪网络传播和交叉压缩单元的高阶特征交互获得准确、全面的物品高阶表示,提高模型推荐精度并增强模型应对数据稀疏场景的能力。在MovieLens-20M、Book-Crossing和Last.FM数据集上的实验结果表明,与KGCN、libFM、CKE等基线方法相比,KGCRN在点击通过率预测、Top-K推荐和应对数据稀疏场景下的性能均得到显著提升,其中,相比KGCN,点击通过率预测实验中KGCRN的AUC增益分别提高0.4、5.1、2.4个百分点,F1值分别提升3.29、2.86、0.96个百分点。
The use of Knowledge Graph(KG)and deep learning for recommendation method has been extensively studied and applied.However,most recommendation models are incomplete in terms of modeling item representation,and the high-order interaction information between users and items has not been fully captured and utilized.A Knowledge Graph Cross Ripple Network(KGCRN)is proposed for the extraction of high-order representations of user and item features,which combines the interaction information between users and items and the association information of KG.The propagation of the ripple network is used to model user preferences and item features,enriching the representation of both,and improving recommendation performance.In addition,an improved cross-compression unit is designed to process the output of the ripple network,utilizing the propagation of the ripple network and the interaction of the high-order features of the cross-compression unit to obtain accurate and comprehensive high-order representations of items,improving the accuracy of model recommendation,and enhancing the ability of the model to deal with sparse data scenarios.Experimental results on the MovieLens-20M,Book-Crossing,and Last.FM datasets show that compared to baseline methods such as KGCN,libFM,and CKE,KGCRN has significantly improved performance in scenarios:Click Through Rate(CTR)prediction,Top-K recommendation,and dealing with data sparsity.Compared to KGCN,the AUC gain in the CTR prediction experiment has increased by 0.4,5.1,and 2.4 percentage points,and the F1 value has increased by 3.29,2.86,and 0.96 percentage points.
作者
唐彦
卢镘旭
TANG Yan;LU Manxu(College of Computer and Information,Hohai University,Nanjing 211100,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2023年第5期63-72,80,共11页
Computer Engineering
基金
国家重点研发计划(2017YFC0405805)。
关键词
知识图谱
深度学习
涟漪网络
改进的交叉压缩单元
推荐系统
Knowledge Graph(KG)
deep learning
ripple network
improved cross-compression unit
recommendation system