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

作　　者：李家乐 王瑞琴[1] 于洋 LI Jiale;WANG Ruiqin;YU Yang(School of Information Engineering,Huzhou University,Huzhou 313000,China)

机构地区：[1]湖州师范学院信息工程学院,浙江湖州313000

出　　处：《电信科学》2024年第11期66-78,共13页Telecommunications Science

基　　金：国家自然科学基金面上项目(No.62277016)。

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。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 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]