基于对比学习的延长增强推荐模型  

Extended reinforcement recommendation model based on contrastive learning

作　　者：张文龙 孙福振[1] 李鹏程 张志伟 王绍卿[1] ZHANG Wenlong;SUN Fuzhen;LI Pengcheng;ZHANG Zhiwei;WANG Shaoqing(School of Computer Science and Technology,Shandong University of Technology,Zibo 255049,China)

机构地区：[1]山东理工大学计算机科学与技术学院,山东淄博255049

出　　处：《山东理工大学学报（自然科学版）》2024年第6期32-38,46,共8页Journal of Shandong University of Technology:Natural Science Edition

基　　金：国家自然科学基金项目(61841602);山东省自然科学基金项目(ZR2020MF147)。

摘　　要：针对现有模型因数据稀疏而导致推荐效果不佳的问题,提出了一种基于对比学习的延长增强推荐模型。首先,通过对交互序列反向延长以获得具有丰富上下文信息的增广序列;其次,为避免现有对比学习正样本生成方式可能会破坏序列语义的问题,将增广序列与原序列进行对比学习以提取两者之间的相似语义;最后,联合训练生成对抗学习任务和对比学习任务以充分挖掘增广序列中丰富的语义特征,从而进行更好推荐。在3个真实数据集上的实验结果表明,所提算法相较于生成对抗网络基线ELECRec和对比学习基线EMKD,评价指标命中率、归一化折损累计增益分别提升了44.3%、43.0%和31.8%、38.7%,验证了所提算法对缓解数据稀疏问题的有效性。In order to solve the problem that the existing model has poor recommendation effect due to sparse data,this paper proposes an extended and enhanced recommendation model based on contrast learning.Firstly,the interaction sequences is inversely extended to obtain augmented sequences with rich contextual information.Secondly,to address the potential issue that existing positive sample generation methods in contrastive learning may disrupt sequence semantics,contrastive learning between the augmented sequences and the original sequences is employed to extract similar semantics between them.Lastly,it jointly trains a generative adversarial learning task and a contrastive learning task to fully explore rich semantic features in the augmented sequences,thereby enhancing recommendations.The experimental results on three real datasets demonstrate that the proposed algorithm has shown a significant improvement of 44.3%,43.0%and 31.8%,38.7%in the evaluation metrics of hit rate and normalized discounted cumulative gain,respectively,compared to the generative adversarial network baseline ELECRec and the contrastive learning baseline EMKD.This validates the effectiveness of the proposed algorithm in alleviating the challenges posed by data sparsity.

关 键 词：推荐模型 对比学习 延长增强 生成对抗网络 

分 类 号：TB399[一般工业技术—材料科学与工程]