DPEKG:基于扩散与自适应去噪增强的知识图谱推荐模型  

DPEKG:Knowledge Graph Recommendation Model Based on Diffusion and Adaptive Denoising Enhancement

作　　者：武征 雒伟群[1,2] Wu Zheng;Luo Weiqun(School of Information Engineering,Xizang Minzu University,Xianyang 712082,China;Xizang Key Laboratory of Optical Information Processing and Visualization Technology,Xianyang 712082,China)

机构地区：[1]西藏民族大学信息工程学院,咸阳712082 [2]西藏光信息处理与可视化技术重点实验室,咸阳712082

出　　处：《西藏科技》2025年第2期62-71,共10页Xizang Science And Technology

基　　金：西藏民族大学校内项目(24MDY07);西藏光信息处理与可视化技术重点实验室开放基金项目(KLFPXZMU2303)。

摘　　要：近年来,知识图谱在推荐系统中为捕捉用户与项目之间的语义关联提供了有效工具。然而,现有研究主要关注简单的直接关系,未能充分利用高阶语义关联,并且在处理图谱噪声时,缺乏对局部结构的精细化操作,冗余信息未能彻底清除,最终影响推荐精度。为解决这些问题,文章提出了一种基于扩散与自适应去噪增强的知识图谱推荐模型DPEKG(Diffusion and Adaptive Post-Enhancement Knowledge Graph)。首先,模型通过路径优化模块,在用户-项目交互图中挖掘多跳路径,构建增强的项目用户交互视图;其次,模型引入扩散去噪机制,逐步扩散并消除知识图谱中的全局噪声,使得经过噪声过滤的图谱能够更加准确地反映真实的用户-项目关联;随后,通过自适应去噪增强模块,模型对去噪后的图谱进行动态加权评分,保留与用户兴趣和项目偏好最相关的节点和边;最后,模型通过对比学习对用户-项目交互视图与去噪后的知识图谱视图进行多视角嵌入对齐,优化推荐效果。实验结果表明,DPEKG在多个公开数据集上的推荐性能显著优于现有方法,验证了其在处理复杂关系和噪声干扰方面的有效性。In recent years,knowledge graphs have provided an effective tool for capturing semantic associations between users and items in recommendation systems.However,existing researches mainly focuses on simple direct relationships and fail to fully utilize higher-order semantic associations,and when dealing with graph noise,it lacks finegrained manipulation of local structures,and redundant information is not thoroughly removed,which ultimately affects recommendation accuracy.To solve these problems,this paper proposes a Diffusion and Adaptive Post-Enhancement Knowledge Graph(DPEKG)recommendation model:first,the model mines multi-hop paths in the useritem interaction graph through the path optimization module to construct an enhanced item-user interaction view;second,the model introduces a diffusion denoising mechanism to gradually diffuse and eliminate the global noise in the knowledge graph,so that the noise-filtered graph can more accurately reflect the real user-item association;subsequently,through the adaptive denoising enhancement module,the model dynamically weights and scores the denoised graph,and retains the nodes and edges most relevant to users’interests and items’preferences;finally,the model performs multi-perspective embedding alignment of the user-item interaction view and the denoised knowledge graph view through contrastive learning to optimize recommendation effects.The experimental results show that the recommendation performance of DPEKG on multiple public datasets significantly outperforms that of existing methods,verifying its effectiveness in dealing with complex relationships and noise interference.

关 键 词：知识图谱 推荐任务 扩散模型 自适应去噪增强 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术] TP391[自动化与计算机技术—计算机科学与技术]