数据增强的多模式时间感知序列推荐
Multi-pattern time-aware sequential recommendation with data augmentation
李家乐 1王瑞琴 1于洋1
作者信息
- 1. 湖州师范学院信息工程学院,浙江 湖州 313000
- 折叠
摘要
序列推荐系统以包含了显式信息的用户交互序列作为上下文,推测用户的下一个可能动作.其中,时间感知序列推荐挖掘了序列中的时间信息,并考虑了时间信息对用户决策的影响.但是现有的时间感知序列推荐模型只运用到了原始时间信息,原本的序列中还有很多额外信息没有被充分挖掘,如用户评分、项目属性、项目流行度以及项目的标题和评论等文本信息.因此,提出了DMTiSASRec模型,它既可以以更高效的方式挖掘时间信息中的相关秩序,还利用对比学习、多模态等技术对不同的额外信息进行挖掘.在5个不同领域、不同规模的公开数据集的实验数据表明,DMTiSASRec比现有模型更有效.
Abstract
In sequential recommendation systems,explicit user interaction sequences are used as context to infer the user's next possible action.Time-aware sequential recommendation models explore the temporal information within the sequence and consider the impact of time on user decisions.However,existing time-aware sequential recommen-dation models only utilize raw temporal information,while many additional pieces of information in the original se-quence are not fully exploited,such as user ratings,item attributes,item popularity,and textual information like item titles and reviews.Therefore,the DMTiSASRec model was proposed,which not only efficiently extracted relevant or-ders beyond temporal information but also leveraged techniques like contrastive learning and multi-modal methods to mine different types of additional information.Experiments on five publicly available datasets across different do-mains and scales show that DMTiSASRec outperforms existing models in terms of effectiveness.
关键词
序列推荐/时间感知/多模态/对比学习Key words
sequential recommendation/time-aware/multi-modal/contrastive learning引用本文复制引用
出版年
2024
NETL
NSTL
万方数据