Journal of Computer Applications ›› 2026, Vol. 46 ›› Issue (7): 2085-2095.DOI: 10.11772/j.issn.1001-9081.2025060795

• Artificial intelligence • Previous Articles    

Knowledge-aware recommendation performance enhancement method based on multi-heterogeneous topology collaboration and hierarchical contrastive learning

Haiwei FAN, Wenshuai DING(), Hongzheng XING   

  1. School of Information Engineering,Chang'an University,Xi'an Shaanxi 710064,China
  • Received:2025-07-17 Revised:2025-09-30 Accepted:2025-10-09 Online:2025-10-15 Published:2026-07-10
  • Contact: Wenshuai DING
  • About author:FAN Haiwei, born in 1974, M. S., associate professor. His research interests include knowledge graph, software system design.
    XING Hongzheng, born in 2002, M. S. candidate. His research interests include knowledge graph, recommender system.

基于多异构拓扑图协同与层次对比学习的知识感知推荐性能增强方法

樊海玮, 丁文帅(), 邢宏政   

  1. 长安大学 信息工程学院,西安 710064
  • 通讯作者: 丁文帅
  • 作者简介:樊海玮(1974—),男,陕西西安人,副教授,硕士,主要研究方向:知识图谱、软件系统设计
    邢宏政(2002—),男,山东聊城人,硕士研究生,主要研究方向:知识图谱、推荐系统。

Abstract:

To address knowledge-aware recommendation limitations, including insufficient high-order semantic modeling, weak exploitation of item deep correlations, and weak exploitation of knowledge graph structure feature, a knowledge-aware recommendation performance enhancement method based on Multi-heterogeneous Topology collaboration and Hierarchical Contrastive Learning (MTHCL) was proposed. First, three heterogeneous graphs were constructed: a Collaborative Filtering-based Item-item Graph (CFIG) was designed to capture collaborative filtering signals, a Knowledge Graph-based Item-entity Graph (KGIG) was established to mine semantic associations, and a User Preference Graph (UPG) was developed to model multi-hop interest evolution via dynamic attention mechanism. Then, a hierarchical contrastive learning strategy was designed to establish a complementary mechanism across heterogeneous topological views. In the local layer, semantic accuracy was ensured by focusing on the direct interaction features between users and items; in the dynamic layer, scene adaptability was enhanced by using attention weighted aggregation of multi-hop relationships; in the global layer, semantic consistency was guaranteed through fusing multi-view information. Finally, regularization term was introduced to suppress overfitting, and an integrated loss function was used to optimize the method. Experimental results demonstrate that MTHCL achieves the highest Area Under Curve (AUC) and F1-score (F1), outperforming the best baseline method — Multi-level Cross-view Contrastive Learning for Knowledge-aware Recommender System (MCCLK) by 2.23 and 2.19 percentage points on the Book-Crossing dataset, 1.24 and 1.78 percentage points on the MovieLens-1M dataset, and 1.86 and 1.97 percentage points on the Last.FM dataset. Ablation study results confirm the critical contributions of multi-view collaboration and hierarchical contrastive learning to performance enhancement, and also verify the superiority of multi-view collaboration and hierarchical contrastive learning in effectively enhancing recommendation performance.

Key words: knowledge-aware recommendation, heterogeneous topology collaboration, hierarchical contrastive learning, dynamic preference modeling

摘要:

针对知识感知推荐中高阶语义关系建模不足、挖掘项目深层关联不充分及知识图谱结构特征利用层次浅的问题,提出一种基于多异构拓扑图协同与层次对比学习的知识感知推荐性能增强方法(MTHCL)。首先,构建3类异构视图:项目-项目视图(CFIG)用于捕捉协同过滤信号,项目-实体视图(KGIG)用于挖掘语义关联,用户偏好图(UPG)则通过动态注意力机制建模多跳兴趣演化;其次,设计分层对比学习策略,与异构拓扑图协同形成互补机制,在局部层聚焦用户-项目的直接交互特征确保基础语义准确性,在动态层利用注意力加权聚合多跳关系提升场景适应性,在全局层融合多视图信息以保证语义一致性;最后,引入正则化项来抑制过拟合,并结合综合损失函数优化方法。实验结果表明, MTHCL的曲线下面积(AUC)和F1分数(F1)均为最高,较最优基线方法——多层次跨视图对比学习知识感知推荐系统(MCCLK)在Book-Crossing数据集上分别提升了2.23和2.19个百分点,在MovieLens-1M数据集上分别提升了1.24和1.78个百分点,在Last.FM数据集上分别提升了1.86和1.97个百分点。消融实验结果也证实了多视图协同与层次化对比学习对性能提升的关键贡献,验证了它们在有效增强推荐性能上的优越性。

关键词: 知识感知推荐, 异构拓扑协同, 层次对比学习, 动态偏好建模

CLC Number: