摘要
为了缓解对比推荐模型因过度依赖结构扰动进行数据增强而导致性能下降的问题,提出知识嵌入增强的对比推荐模型,利用知识图谱的嵌入表征来引导对比学习过程,从而实现高效的物品推荐.通过关系感知的知识聚合模块捕获知识图谱中的异质性关系信息以获得知识嵌入,利用图神经网络编码器从用户-项目交互图中获取实体表征;通过基于知识增强的对比推荐模块将知识嵌入融入用户交互图的表征学习中,强化用户和项目嵌入表示,从而提升推荐精度.在企业服务、书籍和新闻3个数据集上进行大量实验,结果表明,所提模型在处理稀疏数据集时具有明显优势.相较于基线模型KGAT、CKAN,所提模型在Recall和NDCG指标上的平均提升幅度超过20%;与性能优越的KGIN、KGCL、MGDCF等对比学习模型相比,实现了平均10%的性能增益,说明所提方法具有全面的性能优势.
A new contrastive recommendation model was proposed to alleviate the performance degradation caused by the excessive reliance on structural perturbations for data augmentation.The embedding representations of knowledge graphs were leveraged to guide the contrastive learning process for highly effective item recommendation.A relationaware knowledge aggregation module was designed to capture heterogeneous relational information from the knowledge graphs,thereby obtaining knowledge embeddings.A graph neural network encoder was utilized to learn entity representations from the user-item interaction graphs.The knowledge embeddings were incorporated into the representation learning through a knowledge-enhanced contrastive recommendation module,to enhance the user and item embeddings and improve the recommendation accuracy.Extensive experiments were conducted on three datasets of enterprise services,books,and news.The results demonstrated that the proposed model had significant advantages in handling sparse datasets.Compared with the baseline models KGAT and CKAN,the proposed model achieved average improvements of over 20%on the Recall and NDCG metrics.Compared with the state-of-the-art contrastive learning models such as KGIN,KGCL and MGDCF,an average performance gain of 10%was realized.These results demonstrate that the proposed method has comprehensive performance advantages.
作者
谢涛
葛慧丽
陈宁
汪晓锋
李延松
黄晓峰
XIE Tao;GE Huili;CHEN Ning;WANG Xiaofeng;LI Yansong;HUANG Xiaofeng(College of Information Engineering,China Jiliang University,Hangzhou 310018,China;Zhejiang Science and Technology Project Management Service Center,Hangzhou 310006,China;School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《浙江大学学报(工学版)》
北大核心
2026年第1期90-98,共9页
Journal of Zhejiang University(Engineering Science)
基金
国家重点研发计划资助项目(2024YFB3312600)
浙江省“领雁”研发攻关计划资助项目(2024C01107)。
关键词
推荐系统
知识图谱
对比学习
数据增强
数据稀疏
recommendation system
knowledge graph
contrastive learning
data augmentation
data sparsity
