《计算机应用》唯一官方网站 ›› 2024, Vol. 44 ›› Issue (5): 1485-1492.DOI: 10.11772/j.issn.1001-9081.2023050756

• 第十九届中国机器学习会议(CCML 2023) • 上一篇    

无负采样的正样本增强图对比学习推荐方法PAGCL

汪炅1, 唐韬韬1, 贾彩燕1,2()   

  1. 1.北京交通大学 计算机与信息技术学院,北京 100091
    2.交通数据分析与挖掘北京市重点实验室(北京交通大学),北京 100091
  • 收稿日期:2023-06-13 修回日期:2023-07-05 接受日期:2023-07-09 发布日期:2023-08-01 出版日期:2024-05-10
  • 通讯作者: 贾彩燕
  • 作者简介:汪炅(1999—),男,安徽黄山人,硕士研究生,CCF会员,主要研究方向:图神经网络、对比学习、推荐系统
    唐韬韬(1998—),男,湖北安陆人,硕士研究生,CCF会员,主要研究方向:推荐系统、图神经网络
    第一联系人:贾彩燕(1976—),女,宁夏石嘴山人,教授,博士,CCF会员,主要研究方向:机器学习、社会计算、推荐系统。

PAGCL: positive augmentation graph contrastive learning recommendation method without negative sampling

Jiong WANG1, Taotao TANG1, Caiyan JIA1,2()   

  1. 1.School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100091,China
    2.Beijing Key Lab of Traffic Data Analysis and Mining (Beijing Jiaotong University),Beijing 100091,China
  • Received:2023-06-13 Revised:2023-07-05 Accepted:2023-07-09 Online:2023-08-01 Published:2024-05-10
  • Contact: Caiyan JIA
  • About author:WANG Jiong, born in 1999, M. S. candidate. His research interests include graph neural network, contrastive learning, recommendation system.
    TANG Taotao, born in 1998, M. S. candidate. His research interests include recommendation system, graph neural network.

摘要:

对比学习(CL)因能够提取数据本身包含的监督信号而被广泛应用于推荐任务。最近的研究表明,CL在推荐方面的成功依赖于对比损失——互信息噪声对比估计(InfoNCE)损失带来的节点分布的均匀性。此外,另一项研究证明贝叶斯个性化排序(BPR)损失的正项与负项分别带来的对齐性和均匀性有助于提高推荐性能。由于在CL框架中对比损失能够带来比BPR负项更强的均匀性,BPR负项存在的必要性值得商榷。实验分析表明在对比框架中BPR的负项是不必要的,并基于这一观察提出了无需负采样的联合优化损失,可应用于经典的CL方法并达到相同或更高的性能。此外,与专注于提高均匀性的研究不同,为进一步加强对齐性,提出一种新颖的正样本增强的图对比学习方法(PAGCL),该方法使用随机正样本在节点表示层面进行扰动。在多个基准数据集上的实验结果表明,PAGCL在召回率及归一化折损累积增益(NDCG)这两个常用指标上均优于SOTA方法自监督图学习(SGL)、简单图对比学习(SimGCL)等,且相较于基模型轻量化图卷积(LightGCN)的NDCG@20提升最大可达17.6%。

关键词: 推荐系统, 对比学习, 自监督学习, 图神经网络, 数据增强

Abstract:

Contrastive Learning (CL) has been widely used for recommendation because of its ability to extract supervised signals contained in data itself. The recent study shows that the success of CL in recommendation depends on the uniformity of node distribution brought by comparative loss — Infomation Noise Contrastive Estimation (InfoNCE) loss. In addition, the other study proves that Bayesian Personalized Ranking (BPR) loss is beneficial to alignment and uniformity, which contribute to higher recommendation performance. Since the CL loss can bring stronger uniformity than the negative term of BPR, the necessity of the negative term of BPR in CL framework has aroused suspicion. Therefore, this study experimentally disclosed that the negative term of BPR is unnecessary in CL framework for recommendation. Based on this observation, a joint optimization loss without negative sampling was proposed, which could be applied to classical CL-based methods and achieve the same or higher performance. Besides, unlike studies which focus on improving uniformity, a novel Positive Augmentation Graph Contrastive Learning method (PAGCL) was presented, which used random positive samples for perturbation at representation level to further strengthen alignment. Experimental results on several benchmark datasets show that the proposed method is superior to SOTA (State-Of-The-Art) methods like Self-supervised Graph Learning (SGL) and Simple Graph Contrastive Learning (SimGCL) on recall and Normalized Discounted Cumulative Gain (NDCG). The method’s improvement over the base model Light Graph Convolutional Network (LightGCN) can reach up to 17.6% at NDCG@20.

Key words: recommendation system, Contrastive Learning (CL), self-supervised learning, graph neural network, data augmentation

中图分类号: