《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
所属专题: 第十九届中国机器学习会议(CCML 2023)
• 第十九届中国机器学习会议(CCML 2023) • 上一篇 下一篇
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
        
                                                                                                            Jiong WANG1, Taotao TANG1, Caiyan JIA1,2( )
)
			  
			
			
			
                
        
    
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.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%。
中图分类号:
汪炅, 唐韬韬, 贾彩燕. 无负采样的正样本增强图对比学习推荐方法PAGCL[J]. 计算机应用, 2024, 44(5): 1485-1492.
Jiong WANG, Taotao TANG, Caiyan JIA. PAGCL: positive augmentation graph contrastive learning recommendation method without negative sampling[J]. Journal of Computer Applications, 2024, 44(5): 1485-1492.
| 方法 | Yelp2018 | Amazon-Book | Amazon-Kindle | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| LightGCN | 0.063 9 | 0.052 5 | 0.041 0 | 0.031 8 | 0.205 7 | 0.131 5 | 
| LightGCN-WN | 0.031 4 | 0.026 0 | 0.008 3 | 0.006 7 | 0.001 7 | 0.000 9 | 
| SGL | 0.067 5 | 0.055 5 | 0.047 8 | 0.037 9 | 0.209 0 | 0.135 2 | 
| SGL-WN | 0.067 8 | 0.055 9 | 0.048 2 | 0.038 1 | 0.211 2 | 0.136 2 | 
表1 使用不同损失函数的LightGCN和SGL的实验结果
Tab. 1 Experiment results of LightGCN and SGL with different losses
| 方法 | Yelp2018 | Amazon-Book | Amazon-Kindle | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| LightGCN | 0.063 9 | 0.052 5 | 0.041 0 | 0.031 8 | 0.205 7 | 0.131 5 | 
| LightGCN-WN | 0.031 4 | 0.026 0 | 0.008 3 | 0.006 7 | 0.001 7 | 0.000 9 | 
| SGL | 0.067 5 | 0.055 5 | 0.047 8 | 0.037 9 | 0.209 0 | 0.135 2 | 
| SGL-WN | 0.067 8 | 0.055 9 | 0.048 2 | 0.038 1 | 0.211 2 | 0.136 2 | 
| 数据集 | 用户数 | 项目数 | 交互次数 | 密度/% | 
|---|---|---|---|---|
| Yelp2018 | 31 668 | 38 048 | 1 561 406 | 0.130 | 
| Amazon-Kindle | 138 333 | 98 572 | 1 909 965 | 0.014 | 
| Alibaba-iFashion | 300 000 | 81 614 | 1 607 813 | 0.007 | 
表2 实验数据集统计信息
Tab. 2 Statistics of experiment datasets
| 数据集 | 用户数 | 项目数 | 交互次数 | 密度/% | 
|---|---|---|---|---|
| Yelp2018 | 31 668 | 38 048 | 1 561 406 | 0.130 | 
| Amazon-Kindle | 138 333 | 98 572 | 1 909 965 | 0.014 | 
| Alibaba-iFashion | 300 000 | 81 614 | 1 607 813 | 0.007 | 
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | 
|---|---|---|---|
| SGL | |||
| SimGCL | |||
| XSimGCL | |||
| PAGCL | 
表3 不同对比学习方法的最佳超参数
Tab. 3 Best hyperparameters of different constrative learning methods
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | 
|---|---|---|---|
| SGL | |||
| SimGCL | |||
| XSimGCL | |||
| PAGCL | 
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| LightGCN | 0.063 9 | 0.052 5 | 0.205 7 | 0.131 5 | 0.105 3 | 0.050 5 | 
| SGL | 0.067 5(+5.6%) | 0.055 5(+5.7%) | 0.209 0(+1.6%) | 0.135 2(+2.8%) | 0.109 3(+3.8%) | 0.053 1(+5.1%) | 
| NCL | 0.067 0(+4.9%) | 0.056 2(+7.0%) | 0.209 0(+1.6%) | 0.134 8(+2.5%) | 0.108 8(+3.3%) | 0.052 8(+4.6%) | 
| MixGCF | 0.071 3(+11.6%) | 0.058 9(+12.2%) | 0.209 8(+2.0%) | 0.135 5(+3.0%) | 0.108 5(+3.0%) | 0.052 0(+3.0%) | 
| SimGCL | 0.072 1(+12.8%) | 0.060 1(+14.5%) | 0.210 4(+2.3%) | 0.137 4(+4.5%) | 0.115 1(+9.3%) | 0.056 7(+12.3%) | 
| XSimGCL | 0.072 3(+13.1%) | 0.060 4(+15.0%) | 0.214 7(+4.4%) | 0.141 5(+7.6%) | 0.119 6(+13.6%) | 0.058 6(+16.0%) | 
| PAGCL | 0.073 2(+14.6%) | 0.061 2(+16.6%) | 0.217 1(+5.5%) | 0.143 9(+9.4%) | 0.120 8(+14.7%) | 0.059 4(+17.6%) | 
表4 不同SOTA方法在3个基准测试上的总体性能比较
Tab. 4 Overall performance comparison among different SOTA methods on three benchmarks
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| LightGCN | 0.063 9 | 0.052 5 | 0.205 7 | 0.131 5 | 0.105 3 | 0.050 5 | 
| SGL | 0.067 5(+5.6%) | 0.055 5(+5.7%) | 0.209 0(+1.6%) | 0.135 2(+2.8%) | 0.109 3(+3.8%) | 0.053 1(+5.1%) | 
| NCL | 0.067 0(+4.9%) | 0.056 2(+7.0%) | 0.209 0(+1.6%) | 0.134 8(+2.5%) | 0.108 8(+3.3%) | 0.052 8(+4.6%) | 
| MixGCF | 0.071 3(+11.6%) | 0.058 9(+12.2%) | 0.209 8(+2.0%) | 0.135 5(+3.0%) | 0.108 5(+3.0%) | 0.052 0(+3.0%) | 
| SimGCL | 0.072 1(+12.8%) | 0.060 1(+14.5%) | 0.210 4(+2.3%) | 0.137 4(+4.5%) | 0.115 1(+9.3%) | 0.056 7(+12.3%) | 
| XSimGCL | 0.072 3(+13.1%) | 0.060 4(+15.0%) | 0.214 7(+4.4%) | 0.141 5(+7.6%) | 0.119 6(+13.6%) | 0.058 6(+16.0%) | 
| PAGCL | 0.073 2(+14.6%) | 0.061 2(+16.6%) | 0.217 1(+5.5%) | 0.143 9(+9.4%) | 0.120 8(+14.7%) | 0.059 4(+17.6%) | 
| 方法 | 邻接矩阵 | 图编码 | 推荐 | 对比 | 
|---|---|---|---|---|
| SGL | 2Bd | BMd | ||
| SimGCL | 2Bd | BMd | ||
| XSimGCL | 2Bd | BMd | ||
| PAGCL | Bd | BMd | 
表5 各对比方法的时间复杂度
Tab. 5 Time complexities of contrast methods
| 方法 | 邻接矩阵 | 图编码 | 推荐 | 对比 | 
|---|---|---|---|---|
| SGL | 2Bd | BMd | ||
| SimGCL | 2Bd | BMd | ||
| XSimGCL | 2Bd | BMd | ||
| PAGCL | Bd | BMd | 
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| PAGCL-2L | 0.073 2 | 0.061 2 | 0.214 5 | 0.140 5 | 0.120 8 | 0.059 4 | 
| PAGCL-3L | 0.072 6 | 0.060 4 | 0.217 1 | 0.143 9 | 0.120 4 | 0.058 7 | 
表6 PAGCL不同层数L结果比较
Tab. 6 Results of PAGCL with different layers L
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| PAGCL-2L | 0.073 2 | 0.061 2 | 0.214 5 | 0.140 5 | 0.120 8 | 0.059 4 | 
| PAGCL-3L | 0.072 6 | 0.060 4 | 0.217 1 | 0.143 9 | 0.120 4 | 0.058 7 | 
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| LightGCN | 0.063 9 | 0.052 5 | 0.205 7 | 0.131 5 | 0.105 3 | 0.050 5 | 
| LightGCN-WN | 0.031 4 | 0.026 0 | 0.001 7 | 0.000 9 | 0.000 1 | 0.000 0 | 
| SGL | 0.067 5 | 0.055 5 | 0.209 0 | 0.135 2 | 0.109 3 | 0.053 1 | 
| SGL-WN | 0.067 8 | 0.055 9 | 0.211 2 | 0.136 2 | 0.101 8 | 0.050 1 | 
| SimGCL | 0.072 1 | 0.060 1 | 0.210 4 | 0.137 4 | 0.115 1 | 0.056 7 | 
| SimGCL-WN | 0.072 0 | 0.060 1 | 0.211 3 | 0.138 7 | 0.115 1 | 0.056 5 | 
| XSimGCL | 0.072 3 | 0.060 4 | 0.214 7 | 0.141 5 | 0.119 6 | 0.058 6 | 
| XSimGCL-WN | 0.072 6 | 0.060 4 | 0.216 1 | 0.143 2 | 0.118 6 | 0.058 2 | 
| PAGCL-N | 0.072 9 | 0.061 1 | 0.215 5 | 0.140 2 | 0.121 1 | 0.059 3 | 
| PAGCL | 0.073 2 | 0.061 2 | 0.217 1 | 0.143 9 | 0.120 8 | 0.059 4 | 
表7 使用不同损失的SOTA对比学习方法的结果比较
Tab. 7 Result comparison of SOTA CL-based methods with different losses
| 方法 | Yelp2018 | Amazon-Kindle | Alibaba-iFashion | |||
|---|---|---|---|---|---|---|
| R@20 | N@20 | R@20 | N@20 | R@20 | N@20 | |
| LightGCN | 0.063 9 | 0.052 5 | 0.205 7 | 0.131 5 | 0.105 3 | 0.050 5 | 
| LightGCN-WN | 0.031 4 | 0.026 0 | 0.001 7 | 0.000 9 | 0.000 1 | 0.000 0 | 
| SGL | 0.067 5 | 0.055 5 | 0.209 0 | 0.135 2 | 0.109 3 | 0.053 1 | 
| SGL-WN | 0.067 8 | 0.055 9 | 0.211 2 | 0.136 2 | 0.101 8 | 0.050 1 | 
| SimGCL | 0.072 1 | 0.060 1 | 0.210 4 | 0.137 4 | 0.115 1 | 0.056 7 | 
| SimGCL-WN | 0.072 0 | 0.060 1 | 0.211 3 | 0.138 7 | 0.115 1 | 0.056 5 | 
| XSimGCL | 0.072 3 | 0.060 4 | 0.214 7 | 0.141 5 | 0.119 6 | 0.058 6 | 
| XSimGCL-WN | 0.072 6 | 0.060 4 | 0.216 1 | 0.143 2 | 0.118 6 | 0.058 2 | 
| PAGCL-N | 0.072 9 | 0.061 1 | 0.215 5 | 0.140 2 | 0.121 1 | 0.059 3 | 
| PAGCL | 0.073 2 | 0.061 2 | 0.217 1 | 0.143 9 | 0.120 8 | 0.059 4 | 
| 1 | CHEN T, KORNBLITH S, NOROUZI M, et al. A simple framework for contrastive learning of visual representations[C]// Proceedings of the 37th International Conference on Machine Learning. New York: JMLR, 2020: 1597-1607. | 
| 2 | HE K, FAN H, WU Y, et al. Momentum contrast for unsupervised visual representation learning[C]// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Washington, DC: IEEE Computer Society, 2020: 9729-9738. 10.1109/cvpr42600.2020.00975 | 
| 3 | WU J, WANG X, FENG F, et al. Self-supervised graph learning for recommendation[C]// Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2021: 726-735. 10.1145/3404835.3462862 | 
| 4 | YU J, YIN H, XIA X, et al. Are graph augmentations necessary? simple graph contrastive learning for recommendation[C]// Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2022: 1294-1303. 10.1145/3477495.3531937 | 
| 5 | RENDLE S, FREUDENTHALER C, GANTNER Z, et al. BPR: Bayesian personalized ranking from implicit feedback[EB/OL]. [2023-01-12]. . 10.1145/3339363.3339370 | 
| 6 | VAN DEN OORD A, LI Y, VINYALS O. Representation learning with contrastive predictive coding[EB/OL]. [2023-01-06]. . | 
| 7 | WANG C, YU Y, MA W, et al. Towards representation alignment and uniformity in collaborative filtering[C]// Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. New York: ACM, 2022: 1816-1825. 10.1145/3534678.3539253 | 
| 8 | WANG T, ISOLA P. Understanding contrastive representation learning through alignment and uniformity on the hypersphere[C]// Proceedings of the 2020 International Conference on Machine Learning. New York: JMLR, 2020: 9929-9939. | 
| 9 | HUANG T, DONG Y, DING M, et al. MixGCF: an improved training method for graph neural network-based recommender systems[C]// Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. New York: ACM, 2021: 665-674. 10.1145/3447548.3467408 | 
| 10 | PARK D H, CHANG Y. Adversarial sampling and training for semi-supervised information retrieval[C]// Proceedings of the 2019 World Wide Web Conference. New York: ACM, 2019: 1443-1453. 10.1145/3308558.3313416 | 
| 11 | YANG Z, DING M, ZHOU C, et al. Understanding negative sampling in graph representation learning[C]// Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York: ACM, 2020: 1666-1676. 10.1145/3394486.3403218 | 
| 12 | SCARSELLI F, GORI M, TSOI A C, et al. The graph neural network model[J]. IEEE Transactions on Neural Networks, 2009, 20(1): 61-80. 10.1109/tnn.2008.2005605 | 
| 13 | GAO C, WANG X, HE X, et al. Graph neural networks for recommender system[C]// Proceedings of the 15th ACM International Conference on Web Search and Data Mining. New York: ACM, 2022: 1623-1625. 10.1145/3488560.3501396 | 
| 14 | WU S, SUN F, ZHANG W, et al. Graph neural networks in recommender systems: a survey[J]. ACM Computing Survey, 2022, 55(5): No. 97. 10.1145/3535101 | 
| 15 | KIPF T N, WELLING M. Semi-supervised classification with graph convolutional networks[EB/OL]. [2023-01-18]. . 10.48550/arXiv.1609.02907 | 
| 16 | WANG X, HE X, WANG M, et al. Neural graph collaborative filtering[C]// Proceedings of the 42th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2019: 165-174. 10.1145/3331184.3331267 | 
| 17 | HE X, DENG K, WANG X, et al. LightGCN: simplifying and powering graph convolution network for recommendation[C]// Proceedings of the 43th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York: ACM, 2020: 639-648. 10.1145/3397271.3401063 | 
| 18 | VAN DEN BERG R, KIPF T N, WELLING M. Graph convolutional matrix completion[EB/OL]. (2017-06-07) [2023-01-14]. . 10.48550/arXiv.1706.02263 | 
| 19 | YU W, QIN Z. Graph convolutional network for recommendation with low-pass collaborative filters[C]// Proceedings of the 2020 International Conference on Machine Learning. New York: JMLR, 2020: 10936-10945. | 
| 20 | YU J, YIN H, GAO M, et al. Socially-aware self-supervised tri-training for recommendation[C]// Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. New York: ACM, 2021: 2084-2092. 10.1145/3447548.3467340 | 
| 21 | YU J, YIN H, LI J, et al. Self-supervised multi-channel hypergraph convolutional network for social recommendation[C]// Proceedings of the Web Conference 2021. New York: ACM, 2021: 413-424. 10.1145/3442381.3449844 | 
| 22 | YU J, XIA X, CHEN T, et al. XSimGCL: towards extremely simple graph contrastive learning for recommendation[EB/OL]. [2022-12-11]. . 10.1109/tkde.2023.3288135 | 
| 23 | ZHANG J, GAO M, YU J, et al. Double-scale self-supervised hypergraph learning for group recommendation[C]// Proceedings of the 30th ACM International Conference on Information & Knowledge Management. New York: ACM, 2021: 2557-2567. 10.1145/3459637.3482426 | 
| 24 | VAN DER MAATEN L, HINTON G. Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(11): 2579-2605. | 
| 25 | CHEN W, HUANG P, XU J, et al. POG: personalized outfit generation for fashion recommendation at Alibaba iFashion[C]// Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York: ACM, 2019: 2662-2670. 10.1145/3292500.3330652 | 
| 26 | LIN Z, TIAN C, HOU Y, et al. Improving graph collaborative filtering with neighborhood-enriched contrastive learning[C]// Proceedings of the ACM Web Conference 2022. New York: ACM, 2022: 2320-2329. 10.1145/3485447.3512104 | 
| [1] | 杨兴耀, 陈羽, 于炯, 张祖莲, 陈嘉颖, 王东晓. 结合自我特征和对比学习的推荐模型[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2704-2710. | 
| [2] | 杜郁, 朱焱. 构建预训练动态图神经网络预测学术合作行为消失[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2726-2731. | 
| [3] | 唐廷杰, 黄佳进, 秦进. 基于图辅助学习的会话推荐[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2711-2718. | 
| [4] | 杨航, 李汪根, 张根生, 王志格, 开新. 基于图神经网络的多层信息交互融合算法用于会话推荐[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2719-2725. | 
| [5] | 杨莹, 郝晓燕, 于丹, 马垚, 陈永乐. 面向图神经网络模型提取攻击的图数据生成方法[J]. 《计算机应用》唯一官方网站, 2024, 44(8): 2483-2492. | 
| [6] | 唐廷杰, 黄佳进, 秦进, 陆辉. 基于图共现增强多层感知机的会话推荐[J]. 《计算机应用》唯一官方网站, 2024, 44(8): 2357-2364. | 
| [7] | 杨帆, 邹窈, 朱明志, 马振伟, 程大伟, 蒋昌俊. 基于图注意力Transformer神经网络的信用卡欺诈检测模型[J]. 《计算机应用》唯一官方网站, 2024, 44(8): 2634-2642. | 
| [8] | 徐松, 张文博, 王一帆. 基于时空信息的轻量视频显著性目标检测网络[J]. 《计算机应用》唯一官方网站, 2024, 44(7): 2192-2199. | 
| [9] | 蒋小霞, 黄瑞章, 白瑞娜, 任丽娜, 陈艳平. 基于事件表示和对比学习的深度事件聚类方法[J]. 《计算机应用》唯一官方网站, 2024, 44(6): 1734-1742. | 
| [10] | 林欣蕊, 王晓菲, 朱焱. 基于局部扩展社区发现的学术异常引用群体检测[J]. 《计算机应用》唯一官方网站, 2024, 44(6): 1855-1861. | 
| [11] | 韩贵金, 张馨渊, 张文涛, 黄娅. 基于多特征融合的自监督图像配准算法[J]. 《计算机应用》唯一官方网站, 2024, 44(5): 1597-1604. | 
| [12] | 郭洁, 林佳瑜, 梁祖红, 罗孝波, 孙海涛. 基于知识感知和跨层次对比学习的推荐方法[J]. 《计算机应用》唯一官方网站, 2024, 44(4): 1121-1127. | 
| [13] | 黄荣, 宋俊杰, 周树波, 刘浩. 基于自监督视觉Transformer的图像美学质量评价方法[J]. 《计算机应用》唯一官方网站, 2024, 44(4): 1269-1276. | 
| [14] | 徐大鹏, 侯新民. 基于网络结构设计的图神经网络特征选择方法[J]. 《计算机应用》唯一官方网站, 2024, 44(3): 663-670. | 
| [15] | 郭安迪, 贾真, 李天瑞. 基于伪实体数据增强的高精准率医学领域实体关系抽取[J]. 《计算机应用》唯一官方网站, 2024, 44(2): 393-402. | 
| 阅读次数 | ||||||
| 全文 |  | |||||
| 摘要 |  | |||||