In existing contrastive learning methods, most approaches rely on node and edge dropout-based data augmentation to construct contrastive views. However, this often leads to the loss of important information, resulting in suboptimal positive samples. Moreover, the complex architecture of contrastive learning, combined with deeply stacked graph neural networks, frequently causes feature dimension redundancy and increases computational costs. To address these issues, this paper proposes a lightweight contrastive learning recommendation algorithm based on feature optimization and decoupling. The method employs a representation-level data augmentation strategy through cross-layer contrastive learning to generate embeddings and adopts an adaptive penalty mechanism to decouple the correlation between feature dimensions, thereby improving embedding quality and enhancing recommendation performance. First, uniform random noise is added to embeddings at each layer for data augmentation, maximizing mutual information between positive samples. A shared single-pass mechanism is utilized to unify contrastive tasks, achieving a lightweight learning architecture. Second, the user and item representation matrices are subjected to adaptive penalties, dynamically adjusting penalty coefficients at each layer to reduce feature dimension correlations. Experimental results demonstrate the significant effectiveness of the proposed method on the Yelp, Amazon-Books, and Alibaba-iFashion datasets, with improvements in the NDCG and Recall metrics by 8.39%, 8.89%, and 8.3%, and 13.42%, 11.96%, and 16.6%, respectively.
为解决上述问题,本文提出了一个基于特征优化解耦的轻量化对比学习推荐算法(Feature optimization and decoupling-based lightweight contrastive learning for recommendation,FDLCL),该模型通过基于噪声的嵌入增强方法和交叉层对比学习,最大化正样本间互信息并降低计算成本,同时动态地对表示矩阵的特征维度进行相关性惩罚,缓解现有深层堆叠图神经网络产生的特征过相关问题,从而提升嵌入表示的质量。
1 FDLCL模型
1.1 符号定义和描述
本文中,模型输入是用户-项目交互数据,其中:为用户集合,为项目集合,m为用户数量,n为项目数量, R 为用户-物品交互矩阵,为用户物品交互图。
鉴于图神经网络技术在表示学习领域的突出表现,本文采用基于图神经网络的推荐模型对用户-项目二分图进行编码。NSE-LGCN[10]是以LightGCN为骨干改进的一种简单、高效、易于实现的图嵌入方法,可以用于大规模的图表示学习。给定节点表示矩阵和,通过与交互矩阵 R 的矩阵乘法获取用户和项目的结构嵌入。
Multi GCCF通过构建用户-项目交互图、用户-用户图和项目-项目图,结合不同的聚合和转换函数,显式地处理用户和项目之间的高阶信息和相似性,提升了模型嵌入空间表示能力。DGCF论证了将用户复杂的意图解耦能更好地建模用户兴趣偏好并提高了模型的可解释性,迭代地优化用户意图,更高效地提取每个意图的相关信息训练模型。LightGCN通过在用户-项目二分图交互数据上线性传播用户和项目嵌入学习用户偏好信息,去除特征变换和非线性激活模块,提升传统协同过滤模型的性能。在此基础上,NSE-LGCN利用一阶邻接信息构建结构嵌入,在传播过程中,每个节点都可以保持自身的特性,有效地与其他节点区分,缓解过平滑的问题。
WuL, HeX N, WangX, et al. A survey on accuracy-oriented neural recommendation: from collaborative filtering to information-rich recommendation[J]. IEEE Transactions on Knowledge and Data Engineering, 2022, 35(5): 4425-4445.
[2]
ChenG, YuZ, LiN, et al. A survey of graph neural networks for recommender systems: challenges, methods, and directions[J]. ACM Transactions on Recommender Systems, 2023, 1(1): 1-51.
[3]
MunikotiS, AgarwalD, DasL, et al. Challenges and opportunities in deep reinforcement learning with graph neural networks: a comprehensive review of algorithms and applications[J]. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(11): 15051-15071.
CenKe-ting, ShenHua-wei, CaoMiao, et al. A survey on graph contrastive learning[J]. Journal of Chinese Information Processing, 2023, 37(5): 1-21.
[6]
WuJian-can, WangXiang, FengFu-li, et al. Self-supervised graph learning for recommendation[C]∥The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Online, Canada, 2021: 726-735.
[7]
ZhouX, LinD H, LiuY, et al. Layer-refined graph convolutional networks for recommendation[C]∥The 2023 IEEE 39th International Conference on Data Engineering, Anaheim, United States, 2023: 1247-1259.
[8]
LiC L, XiaL H, RenX B, et al. Graph transformer for recommendation[C]∥The 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, Taiwan, China, 2023: 1680-1689.
[9]
XiaL H, HuangC, XuY, et al. Hypergraph contrastive collaborative filtering[C]∥The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Taiwan, China, 2022: 70-79.
[10]
JinW, LiuX R, MaY, et al. Feature overcorrelation in deep graph neural networks: a new perspective[C]∥The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, USA, 2022: 709-719.
[11]
JinX Z, LiJ T, XieY Z, et al. Enhancing graph collaborative filtering via neighborhood structure embedding[C]∥The 2023 IEEE International Conference on Data Mining, Shanghai, China, 2023: 190-199.
[12]
RendleS, FreudenthalerC, GantnerZ, et al. BPR: bayesian personalized ranking from implicit feedback[C]∥The 25th Conference on Uncertainty in Artificial Intelligence, Montreal, Canada, 2009: 452-461.
[13]
RenMeng-wei, DelbracioM, TalebiH, et al. Multiscale structure guided diffusion for image deblurring[C]∥IEEE/CVF International Conference on Computer Vision, Paris, France, 2023: 10687-10699.
[14]
YuJ L, XiaX, ChenT, et al. XSimGCL: towards extremely simple graph contrastive learning for recommendation[J]. IEEE Transactions on Knowledge and Data Engineering, 2023, 36(2): 913-926.
[15]
WuW, WangC, ShenD Z, et al. AFDGCF: adaptive feature de-correlation graph collaborative filtering for recommendations[C]∥The 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, Washington, USA, 2024: 1242-1252.
[16]
FanW Q, LiuX R, JinW, et al. Graph trend filtering networks for recommendation[C]∥The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, 2022: 112-121.
[17]
WangY, ZhaoY Y, ZhangY, et al. Collaboration-aware graph convolutional network for recommender systems[C]∥Proceedings of the ACM Web Conference 2023, Austin, USA, 2023: 91-101.
[18]
HuangT L, DongY X, DingM, et al. Mixgcf: an improved training method for graph neural network-based recommender systems[C]∥The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, 2021: 665-674.
[19]
WangX, JinH Y, ZhangA, et al. Disentangled graph collaborative filtering[C]∥The 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, China, 2020: 1001-1010.