基于协作知识图谱的个性化推荐模型  

Personalized recommendation model based on collaborative knowledge graph

作　　者：李军[1] LI Jun(School of Computer and Information Technology,Anhui Vocational and Technical College,Hefei Anhui 230011)

机构地区：[1]安徽职业技术学院计算机与信息技术学院,安徽合肥230011

出　　处：《宁夏师范大学学报》2025年第1期91-99,共9页Journal of Ningxia Normal University

摘　　要：针对知识图谱关系类型的多样性及实体间的语义关系难以准确捕捉的问题,提出基于协作知识图谱的个性化推荐模型.在保留用户-项目交互和知识图谱全局结构路径信息的基础上,利用全局路径聚合模块获取用户嵌入和项目嵌入,在知识图谱上利用用户偏好感知模块提取项目嵌入.然后集成用户-项目交互和知识图谱部分的项目嵌入得到最终的项目嵌入,再与用户嵌入进行内积操作得到最后的预测评分.实验结果表明,该模型在点击率预测场景AUC和F1指标优于几种先进的对比模型,尤其是在Last.FM数据集的F1指标提升了1.37%,且运行时间较为合理.消融实验表明,全局路径聚合模块和用户偏好感知模块有助于提升个性化推荐模型的精度.Considering the diversity of relationship types of knowledge graph and the difficulty of accurately capturing semantic relationships between entities,a personalized recommendation model based on collaborative knowledge graph is proposed.On the basis of retaining the user-item interaction and the global structural path information of the knowledge graph,the global path aggregation module is utilized to obtain user embeddings and item embeddings,and then the item embeddings are extracted on the knowledge graph using the user preference-aware module.Next,the final item embeddings are obtained by integrating the item embeddings in the user-item interaction and knowledge graph parts,and the final prediction scores are derived from the inner product operation with the user embeddings.The experimental results show that the model outperforms several state-of-the-art comparison models in the click-through rate prediction scenario for AUC and F1 metrics,especially the F1 metrics under the Last.FM dataset improves by 1.37%,and the runtime is reasonable.The ablation experiments show that the global path aggregation module and the user preference-aware module may help to improve the accuracy of the personalized recommendation model.

关 键 词：推荐系统 知识图谱 图神经网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]