基于图神经网络的挖掘潜在偏好图推荐算法

Recommendation Algorithm for Mining Potential Preference Graph Based on Graph Neural Network

知识图谱在推荐系统中扮演着越来越重要的角色,最新技术趋势是开发基于图神经网络的端到端推荐模型。然而,现有基于GNN的模型通常未能充分挖掘知识图谱中的信息,仅简单地将用户通过项目与知识图谱中的实体相连,未明确建模用户与实体之间的关系。为此,提出一种基于图神经网络的挖掘潜在偏好图的推荐算法UEKR,从协同知识图谱中动态提取用户感兴趣的实体,并建模用户与实体之间的关系,构建用户—实体关系图,以丰富用户表示,增强推荐性能。在3个基准数据集上的实验结果表明,UEKR相较对照模型在AUC指标方面提升了0.75%~3.65%,在F1指标方面提升了0.70%~1.75%。

Graph recognition plays an increasingly important role in recommendation systems,and the latest technological trend is to develop end-to-end recommendation models based on graph neural networks.However,existing GNN based models often fail to fully explore the information in the knowledge graph,simply connecting users to entities in the knowledge graph through projects,without clearly modeling the relationships between users and entities.To this end,a recommendation algorithm UEKR based on graph neural networks is proposed for mining latent preference maps.It dynamically extracts entities of interest to users from collaborative knowledge graphs,models the relationship between users and entities,and constructs a user entity relationship graph to enrich user representation and enhance recommendation performance.The experimental results on three benchmark datasets showed that UEKR improved AUC indicators by 0.75%to 3.65%and F1 indicators by 0.70%to 1.75%compared to the control model.