Abstract: Under conditions of data sparsity and noisy scenarios, existing graph neural network-based collaborative filtering models face multiple challenges: static noise injection obscures true signals, fixed semantic prototypes fail to capture dynamic user interests, and complex augmentation methods incur substantial computational overhead. To address these issues, proposed methodology introduces contrastive collaborative filtering integrating graph diffusion generation with adaptive sampling. Specifically, lightweight diffusion modeling employing progressive denoising optimizes node representations via forward noising and reverse denoising processes, generating robust contrastive views resistant to noise. Furthermore, combining random masking with restart walk algorithms enables collaborative modeling of local neighborhood features alongside global structural semantics, yielding high-quality negative samples. Optimization employs modified InfoNCE(Information Noise Contrastive Estimation) loss function for multi-view contrastive learning objectives, enhancing representation discriminability. Evaluations on Gowalla, Yelp, and Amazon-Book datasets demonstrate performance improvements: Recall@20 increased by 0.63%, 2.88%, and 1.88% respectively against strongest baselines; NDCG@40 improved by 1.26%, 1.84%, and 12.51%; long-tail user recommendation performance elevated by 26.7%; training efficiency enhanced by 90% with convergence speed exceeding contrastive methods by 32%. These advancements significantly improve noise robustness and dynamic adaptability for recommendation systems in open environments.

Key words: Keywords: graph neural networks, contrastive learning, self-supervised learning, diffusion models, graph algorithms

摘要： 针对现有基于图神经网络的协同过滤方法在数据稀疏和噪声场景下存在的静态噪声注入易掩盖真实信号、固定语义原型难以捕捉动态兴趣、复杂增强方法计算开销大等问题，提出一种基于图扩散生成与自适应采样的对比协同过滤方法。首先设计基于渐进去噪的轻量级扩散模型，通过前向加噪和反向去噪优化节点表示，生成抗噪性强的对比视图；其次结合随机掩码与重启游走算法，协同建模局部邻域特征与全局结构语义，生成高质量负样本；最后通过改进的InfoNCE（Information Noise Contrastive Estimation）损失函数优化多视图对比学习目标，提升表征判别性。在Gowalla/Yelp/Amazon-Book数据集上，Recall@20指标较最优基线提升0.63%/2.88%/1.88%，NDCG@40指标提升1.26%/1.84%/12.51%，长尾用户推荐效果相对提升26.7%，训练效率提升90%且收敛速度优于对比方法32%，显著提升了开放环境下推荐系统的抗噪性与动态适应性。

关键词: 关键词: 图神经网络, 对比学习, 自建督学习, 扩散模型, 图算法