| Uni-Rec：基于生成式扩散与协同知识图谱卷积的推荐算法 |
投稿时间：2026-05-12 修订日期：2026-06-02  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 45 |
| 全文下载次数: 0 |
|
| 基金项目:国网湖南电力科技项目,全链路数据血缘追踪关键技术研究项目(5216A624000L) |
|
| 中文摘要:在推荐系统领域,数据稀疏性与知识图谱(KG)固有的结构噪声是限制模型性能的两大瓶颈。现有方法多依赖隐式对比学习或单模态扩散,难以实现显式结构净化与多模态语义对齐的联合建模。本文提出了一种名为Uni-Rec的生成式扩散推荐框架。该框架创新性地结合了图净化技术与扩散生成范式：首先,设计了结构感知编码器,通过互信息剪枝与协同知识图谱卷积(CKGC)从去噪后的异构图中提取鲁棒的节点表征；其次,通过生成式联合扩散模块,在潜在空间重构用户-物品的联合分布,并引入多模态信号注入(MSI)增强语义对齐；最后,利用向量化检索器与加速采样策略实现高效推理。在Yelp2018、Amazon-Book和Alibaba-iFashion三个基准数据集上的实验表明,Uni-Rec在Recall和NDCG指标上均显著优于现有扩散与知识图谱推荐模型。 |
| 中文关键词:推荐系统 知识图谱 生成式扩散模型 协同知识图谱卷积 多模态融合 图卷积网络 图剪枝 |
| |
| Uni-Rec: A Recommendation Algorithm Based on Generative Diffusion and Collaborative Knowledge Graph Convolution |
|
|
| Abstract:In the field of recommendation systems, data sparsity and the inherent structural noise of knowledge graphs (KG) are two major bottlenecks limiting model performance. Existing methods mostly rely on implicit contrastive learning or single-modal diffusion, making it difficult to achieve joint modeling of explicit structural purification and multi-modal semantic alignment. This paper proposes a generative diffusion recommendation framework named Uni-Rec. The framework innovatively integrates graph purification technology with the diffusion generation paradigm: first, a structure-aware encoder is designed to extract robust node representations from the denoised heterogeneous graph via mutual information pruning and Collaborative Knowledge Graph Convolution (CKGC); second, a generative joint diffusion module reconstructs the user-item joint distribution in the latent space and introduces Multi-modal Signal Injection (MSI) to enhance semantic alignment; finally, efficient inference is achieved through a vectorized retriever and an accelerated sampling strategy. Experiments on three benchmark datasets—Yelp2018, Amazon-Book, and Alibaba-iFashion—demonstrate that Uni-Rec significantly outperforms existing diffusion-based and knowledge-graph-enhanced recommendation models in both Recall and NDCG metrics. |
| keywords:recommendation system knowledge graph generative diffusion model collaborative knowledge graph convolution multi-modal fusion graph convolutional network graph pruning |
| 查看全文 查看/发表评论 下载pdf阅读器 |
位访问者,今日访问
