融合成对损失函数与分级图卷积网络的协同排名模型  

Collaborative ranking model based on hierarchical graph convolution network with pair-wise loss

作　　者：郑升旻 胡林发 漆鑫鑫 ZHENG Shengmin;HU Linfa;QI Xinxin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650504

出　　处：《现代电子技术》2024年第4期169-175,共7页Modern Electronics Technique

摘　　要：协同排名方法是一类直接优化推荐项目序列的推荐方法,但在显式评分场景中,现有的协同排名算法对用户和项目间的交互信息利用不足,使用的交互函数对用户和项目间的非线性交互关系建模不充分。针对此问题,提出一种融合成对损失函数与分级图卷积网络的协同排名模型。首先根据评分数据构造用户和项目的分级异质交互二部图以及项目间的成对比较集合;之后利用分级图卷积网络挖掘用户和项目间的异质交互关系,并设计相应的自连接操作;接着利用神经网络融合辅助信息以构造两者的嵌入表示,结合神经网络与内积构造用户项目间的交互函数以建模非线性关系;最后基于成对比较数据优化模型,提升模型预测排名质量。在多个公开数据集上与同类方法的对比实验结果表明,所提算法预测排名质量较优。Collaborative ranking methods are a class of recommendation methods that directly optimize the sequence of recommended items,but the existing collaborative ranking algorithms do not make sufficient use of the interaction information between users and items in the explicit scoring scenario,and the interaction functions used do not adequately model the nonlinear interaction between users and items.Therefore,a collaborative ranking model based on hierarchical graph convolution network with pair-wise loss is proposed to address this problem.The heterogeneous interaction bipartite graphs of users and items and pairwise comparison sets are first constructed based on the rating data.The hierarchical convolutional network with specific self-connection operation is used to mine the heterogeneous interactions between users and items and use neural network to fuse the auxiliary information,so as to construct both embedding representation.Then neural network and inner product are combined to construct the interaction function between users and items to modelling the nonlinear relationship.This model is optimized based on pairwise comparisons to improve the quality of ranking it predicting.The comparison experiments with similar methods on several public datasets show that the prediction ranking quality of the propsoed algorithm is better.

关 键 词：协同排名 成对损失函数 图卷积神经网络 异质交互图 自连接 非线性关系 

分 类 号：TN926-34[电子电信—通信与信息系统] TP391.3[电子电信—信息与通信工程]