摘要
为了缓解目前基于动态图表征的序列推荐研究中存在的用户交互数据高稀疏且包含大量噪音、模型训练需要大量有标记样本等问题,本文提出了基于对比学习的动态图序列推荐方法(CDGSR).具体而言,CDGSR从粗粒度和细粒度不同层面设计了层间对比学习、两次传播对比学习和随机噪声扰动对比学习三种不同角度的对比学习方法.实验表明,CDGSR在Amazon-Beauty、Amazon-Games、Amazon-CDs三个现实数据集上的归一化折损累计增益NDCG@10分别达到了0.363 3、0.587 3、0.522 0,Hit@10分别达到了0.525 8、0.778 6、0.735 9.与基于矩阵分解的BPR-MF、FPMC,基于神经网络的GRU4Rec、Caser、SASRec和基于图神经网络的SR-GNN、HGN、Hyper-Rec、DGSR等方法相比,CDGSR均取得了最好的结果.其中,与性能最好的DGSR相比,CDGSR在Amazon-CDs数据集上的NDCG@10提升了1.97%,Hit@10提升了1.60%.这些结果表明,本文提出的CDGSR能够有效利用对比学习方法提升动态图序列推荐方法的推荐性能.
Abstract
To alleviate the problems in dynamic graph sequential recommendation,such as sparse and noisy user-item interaction da-ta,and the requirement for a large number of labels,this paper proposes a new dynamic graph sequential recommendation method based on contrastive learning,which is called CDGSR(Contrastive Dynamic Graph for Sequential Recommendation).Specifically,CDGSR designed three different contrastive learning methods from coarse-grained to fine-grained:inter layer contrastive learning,twice propagation contrastive learning and random noise perturbation contrastive learning.The experimental results demonstrate that CDGSR achieves NDCG@10 scores of 0.363 3,0.587 3,and 0.522 0 on the real-world datasets of Amazon-Beauty,Amazon-Games,and Amazon-CDs,respectively.Additionally,the corresponding Hit@10 scores are 0.525 8,0.778 6,and 0.735 9.Compared to ma-trix factorization-based methods like BPR-MF and FPMC,neural network-based methods like GRU4Rec,Caser,SASRec,and graph neural network-based methods like SR-GNN,HGN,HyperRec,and DGSR,CDGSR consistently achieves the best results.Specifi-cally,compared to the best-performing method DGSR,CDGSR improves NDCG@10 by 1.97%and Hit@10 by 1.60%on the Ama-zon-CDs dataset.These results indicate that CDGSR can effectively utilize contrastive learning to improve the performance of dy-namic graph sequential recommendation method.
基金项目
浙江大学上海高等研究院繁星科学基金(SN-ZJU-SIAS-001)
国家自然科学基金(62372399)
国家科技期刊平台
NSTL
万方数据