基于自适应增强的多视图对比推荐算法  

Multi-view Contrastive Recommendation Algorithm Based on Adaptive Enhancement

作　　者：姚迅[1] 王海鹏 胡新荣[1] 杨捷 YAO Xun;WANG Haipeng;HU Xinrong;YANG Jie(School of Computer Science and Artificial Intelligence,Wuhan Textile University,Wuhan 430200,Hubei,China;Faculty of Engineering and Information Sciences,University of Wollongong,Wollongong 2259,Australia)

机构地区：[1]武汉纺织大学计算机与人工智能学院,湖北武汉430200 [2]伍伦贡大学工程与信息科学学院,澳大利亚伍伦贡2259

出　　处：《计算机工程》2025年第5期103-113,共11页Computer Engineering

摘　　要：近年来,基于神经网络架构的推荐系统取得了显著成功,但在处理富含流行偏见和交互噪声的数据时,未能达到期望的效果。对比学习作为一种从无标记数据中学习的新兴技术备受关注,为解决这一问题提供了潜在方案。提出一种端到端的图对比推荐算法AMV-CL。首先,基于节点的潜在表征构建用户-项目交互图的互补图;其次,引入自适应增强技术,分别从节点和边缘角度生成多视图数据,并通过重参数化网络调整图结构;最后,规范化对比损失中锚节点的正样本来源,同时利用多视图对比损失来学习用户/项目的潜在表征。在公共数据集上的实验结果显示,相较于最优基准方法SimGCL,AMV-CL在评价指标Recall@20和NDCG@20上的提升最高可达到12.03%和12.64%,表明所提方法能够有效提升推荐性能。Recommendation systems based on neural network architectures have achieved remarkable success in recent years;however,they fail to achieve the desired results when dealing with data rich in popularity biases and interaction noise.Contrastive learning,an emerging technology for learning from unlabeled data,has attracted considerable attention and provides a potential solution to this problem.This study proposes an end-to-end graph-contrastive recommendation method called AMV-CL.This method first constructs a complementary graph of a user-item interaction graph based on the latent representation of nodes and then introduces adaptive augmentation to generate multi-view data from node and edge perspectives.Subsequently,it adjusts the graph structure through a reparameterization network and finally normalizes the sources of positive samples of anchor nodes in contrastive loss,while leveraging multi-view contrastive loss to learn latent representations of users/items.A large number of experiments on public datasets show that,compared with the optimal benchmark method SimGCL,AMV-CL yields up to 12.03%and 12.64%improvements in the Recall@20 and NDCG@20 evaluation indicators,respectively.Experimental results show that the proposed method can effectively improve the recommendation performance.

关 键 词：图神经网络 推荐系统 多视图 对比学习 自适应增强 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]