基于对比学习和元优化学习的序列推荐方法  

Sequential Recommendation Method Based on Contrastive Learning and Meta-optimized Learning

作　　者：谢林泽 陈平华[1] 邓柏城[2] XIE Lin-ze;CHEN Ping-hua;DENG Bai-cheng(School of Computer Science,Guangdong University of Technology,Guangzhou 510006,China;School of Economics and Management,Guangdong Institute of Science and Technology,Zhuhai 519090,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006 [2]广东科学技术职业学院经济管理学院,广东珠海519090

出　　处：《计算机技术与发展》2024年第10期148-155,共8页Computer Technology and Development

基　　金：广东省重点领域研发计划项目(2023B1111050010,2020B0101100001)。

摘　　要：序列推荐是根据用户和项目的历史交互记录对用户兴趣建模,进行下一项目推荐。对比学习(CL)作为一种辅助信息能够有效地提高推荐模型质量,但现有基于对比学习的序列推荐方法采取随机数据增强方式存在的效果不稳定及难以泛化的问题,为此,提出了一种基于对比学习和元优化学习的序列推荐方法。首先,在数据增强环节,根据序列中项目之间的时间间隔为序列生成数据分布更加均匀的数据增强视图;其次,构建可学习的模型增强模块,用于捕获数据增强视图中潜在的语义信息,增强模型的泛化能力;最后,为解决数据增强模块和模型增强模块之间不同优化目标问题,使用元优化学习方法优化更新两个模块之间的参数,进而完成推荐。在Beauty、Sports和Yelp等三个公开数据集上的实验结果显示,在召回率和归一化折损累计增益指标上,相较于其它基线模型,CLMLRec均有显著提升,表明该模型具有良好的推荐性能。Sequential recommendation is to model user interest based on historical interaction records between users and projects,and recommend the next project.As a kind of side information,contrast learning(CL)can effectively improve the quality of recommendation models,but the existing sequential recommendation methods based on contrast learning are unstable and difficult to generalize by using random data enhancement.To address the above problem,a sequence recommendation method based on contrastive learning and meta-optimized learning is proposed.Firstly,in the data augmentation step,the data augmentation view with more uniform data distribution is generated for the sequence according to the time interval between the items in the sequence.Secondly,a learnable model augmentation module is constructed to capture the potential semantic information in the data augmentation view and enhance the generalization ability of the model.Finally,in order to solve the problem of different optimization objectives between the data augmentation module and the model augmentation module,the meta-optimized learning is used to optimize and update the parameters between the two modules to complete the recommendation.Experimental results on three publicly available datasets,including Beauty,Sports and Yelp,showed that CLMLRec has significantly improved in terms of recall and NDCG compared with other baseline models,indicating that the model has good recommendation performance.

关 键 词：序列推荐 对比学习 元优化学习 数据增强 模型增强 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]