融合时间感知和多兴趣提取网络的序列推荐
Fusing time-aware and multi-interest extraction network for sequential recommendation
唐宏 1金哲正 1张静 1刘斌1
作者信息
- 1. 重庆邮电大学 通信与信息工程学院,重庆 400065;重庆邮电大学 移动通信技术重庆市重点实验室,重庆 400065
- 折叠
摘要
针对序列推荐任务中的时间动态性和多重兴趣建模问题,提出一种时间感知的项目嵌入方法,用于学习项目之间的时间关联性.在此基础上,提出一种融合时间感知和多兴趣提取网络的序列推荐(time-aware multi-inter-est sequence recommendation,TMISA)方法.TMISA采用自注意力序列推荐模型作为局部特征学习模块,以捕捉用户行为序列中的动态偏好;通过多兴趣提取网络对用户的全局偏好进行建模;引入门控聚合模块将局部和全局特征表示动态融合,生成最终的用户偏好表示.实验证明,在 5 个真实推荐数据集上,TMISA模型表现出卓越性能,超越了多个先进的基线模型.
Abstract
To address the challenges of temporal dynamics and multi-interest modeling in sequence recommendation tasks,this paper proposes a time-aware item embedding method to learn the temporal correlations among items.Building upon this,we introduce a sequence recommendation approach called time-aware multi-interest sequence recommendation(TMI-SA),which integrates time awareness with multi-interest extraction networks.Firstly,TMISA employs a self-attention-based sequence recommendation model as a local feature learning module to capture dynamic user preferences within behav-ior sequences.Secondly,it models global user preferences through a multi-interest extraction network.Finally,a gate aggre-gation module dynamically merges local and global feature representations to generate the ultimate user preference represen-tation.Experimental results demonstrate that the TMISA model exhibits outstanding performance across five real-world rec-ommendation datasets,surpassing multiple state-of-the-art baseline models.
关键词
序列推荐/自注意力机制/时间感知的项目嵌入/多兴趣提取网络/门控聚合模块Key words
sequential recommendation/self-attention/time-aware item embedding/multi-interest extraction network/ga-ted aggregation module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