基于虚拟关系知识图可自适应聚合的推荐算法
Recommendation Algorithm Based on Adaptive Aggregation of Virtual Relationship Knowledge Graph
李源 1杨谋均2
作者信息
- 1. 湖南省电子信息产业研究院,湖南长沙 410001
- 2. 中车株洲电力机车研究所有限公司,湖南株洲 412001
- 折叠
摘要
在信息爆炸的时代,推荐算法成为应对信息过载的有效手段.近年来,图神经网络(GNN)以其强大的建模能力和应对冷启动的优势被广泛应用于推荐算法.本文提出了一种基于深度强化学习与GNN-R的联合训练框架,解决GNN-R中固定层数和聚合策略的问题,通过间隔经验回放和延后奖励机制,优化了推荐模型的学习过程.在此基础上,提出了自适应优化GNN-R聚合层数和虚拟关系数量的两个优化算法,改进了 VRKG4Rec模型的性能.实验结果表明,两个优化算法对比VRKG4Rec模型都有较好的性能提升.
Abstract
In the era of information explosion,recommendation algorithms have become an effective means to cope with information overload.In recent years,graph neural networks(GNN)have been widely applied in recommendation algo-rithms due to their powerful modeling capabilities and advantages in addressing cold start issues.A joint training framework based on deep reinforcement learning and GNN-R is proposed in this paper to address the fixed-layer and aggregation strate-gy issues in GNN-R.By employing interval experience replay and delayed reward mechanisms,the learning process of the recommendation model is optimized.Building upon this,two optimization algorithms for adaptively optimizing the aggrega-tion layers and virtual relation quantities in GNN-R are proposed,enhancing the performance of the VRKG4Rec model.Ex-perimental results demonstrate significant performance improvements of the two optimization algorithms compared to the VRKG4Rec model.
关键词
推荐算法/图神经网络/深度强化学习/知识图谱Key words
recommendation algorithm/graph neural network/deep reinforcement learning/knowledge graph引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据