Abstract: A knowledge-aware recommendation model, called Fusion of Denoising Strategies and Multi-View Contras-tive Learning (FDSMVC), is proposed to address the issues of poor noise reduction effect, inadequate extrac-tion of semantic information between items, and imbalanced utilization of heterogeneous information in knowledge graph-based recommendation algorithms. Firstly, noise reduction is performed on the user-item interaction graph and the knowledge graph through selectively dropping edges and masking low-weight triplets using a weighted function Secondly, random singular value decomposition, cosine similarity, k-nearest neighbors (kNN) sparsity, and path-based graph attention networks are used to construct collaborative view, semantic view, and structural view of item-item relationships. Next, contrastive learning is applied to learn representations within each view, considering intra-graph, local, and global contexts. Finally, a multi-task optimization framework is applied to jointly optimize the recommendation task and the contrastive learning task, resulting in the prediction of user-item interactions. Experimental evaluations on five publicly available datasets demonstrate that the proposed FDSMVC model outperforms the state-of-the-art baseline methods, with improvements of 1.06%-2.04% in terms of AUC, 1.52%-2.60% in terms of F1 score, and better recall@K performance compared to the best baseline method.

Key words: Recommendation system, Knowledge graph, Contrastive learning, Graph neural network, Self-supervised learning

摘要： 针对现有基于知识图谱的推荐算法中存在的降噪效果不佳、项目间语义信息提取不足、信息利用不平衡问题，提出一种融合降噪策略与多视图对比学习(Fusion of denoising strategies and multi-view contrastive learning，FDSMVC)的知识感知推荐模型。首先分别以选择性丢边和加权函数掩盖低权重三元组的方式对用户项目交互图与知识图进行降噪；其次分别采用随机奇异值分解法、余弦相似度与k-最近邻(k-nearest neighbors, kNN)稀疏法和基于路径的图注意力网络构建协同视图、项目间的语义视图和结构视图；接下来将多个视图进行图内、局部和全局三种对比学习；最后利用多任务策略对推荐任务和对比学习任务进行联合优化得到用户与项目交互的可能性。在五个真实的公开数据集上进行了实验，相较于性能最优的基线方法，FDSMVC模型在AUC和F1指标分别提升了1.06%-2.04%和1.52%-2.60%，recall@K指标也优于最佳基线。

关键词: 推荐系统, 知识图谱, 对比学习, 图神经网络, 自监督学习