Abstract: In Graph Neural Network (GNN)-based recommendation systems, traditional neighbor aggregation mechanisms are difficult to capture personalized user-item preferences and limited the modeling of complex interactions; random-strategy negative sampling is prone to generate low-difficulty negative samples, leading to blurred decision boundaries and compromised training efficiency and recommendation performance. To address these issues, a graph recommendation model DEMS (Dynamic Embedding and Mixup Sampling model for graph recommendation), integrating dynamic embedding enhancement and hybrid sampling, was proposed. Firstly, a dynamic embedding enhancement module was introduced in the encoding stage to mine fine-grained personalized user preferences; secondly, a dimension-decoupled Region Mixed Sampling （RMS） strategy was applied in training to generate moderate-difficulty pseudo-negative samples via interpolation; finally, a dynamic screening mechanism was designed using positive sample prediction scores to adjust negative sampling distribution, optimizing sample quality and model convergence to improve overall recommendation effect. Experimental results on datasets with different sparsity show that compared with mainstream baseline models such as LEGCF (Lightweight Embeddings for Graph Collaborative Filtering), DEMS achieves 6.6% and 7.7% improvements in Recall@40 (Recall) and NDCG@40 (Normalized Discounted Cumulative Gain) for the Top-40 recommendation list, respectively, verifying the model’s effectiveness in recommendation tasks.

Key words: Graph Neural Network (GNN), embedding enhancement, negative sample mining, mixup sampling strategy, personalized recommendation

摘要： 图神经网络(GNN)推荐系统中，传统的邻居聚合机制难以捕捉用户与项目之间的个性化偏好,限制了对复杂交互关系的建模能力；基于随机策略的负采样方法易生成训练难度较低的负样本，导致决策边界模糊进而影响模型训练效率和推荐效果。针对上述问题，提出一种基于动态嵌入增强与混合采样融合(DEMS)的图推荐模型。首先，在编码阶段引入动态嵌入增强模块，细粒度地挖掘用户的个性化偏好特征；其次，在训练阶段运用维度解耦的区域混合采样(RMS)策略，基于插值机制生成难度适中的伪负样本;最后，结合正样本的预测得分设计动态筛选机制调整负采样分布，进一步优化训练过程的样本质量与模型收敛效率，从而提升整体推荐效果。在不同稀疏度数据集上的实验结果表明，相较于主流的LEGCF(Lightweight Embeddings for Graph Collaborative Filtering)等基线模型，DEMS在推荐列表Top-40对应的召回率（Recall@40）和归一化折损累积增益（NDCG@40）上，分别实现6.6%和7.7%的提升，验证了该模型在推荐任务中的有效性。

关键词: 图神经网络, 嵌入增强, 负样本挖掘, 混合采样策略, 个性化推荐