基于对比增强时间感知自注意力机制的序列推荐  

Sequential recommendation based on contrast enhanced time-aware self-attention mechanism

作　　者：于洋 王瑞琴[1] YU Yang;WANG Ruiqin(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》2025年第1期137-147,共11页Telecommunications Science

基　　金：国家自然科学基金面上项目(No.62277016)。

摘　　要：现有序列推荐模型在绝对交互时间的利用上存在不足,导致用户偏好建模不准确。因此,提出了基于对比增强时间感知自注意力机制的序列推荐模型(sequential recommendation based on contrast enhanced timeaware self-attention mechanism,CTiSASRec)。首先,注意力权重的计算过程整合了评分数据、绝对交互时间、位置信息和项目流行度;其次,将项目的绝对交互时间和位置顺序融合,生成新的项目位置嵌入;最后,训练过程中利用对序列两次建模结果的对比学习来区分样本间的相似性和差异性,进而提高模型的准确性和鲁棒性。在6个不同领域和规模的数据集上进行的实验表明,CTiSASRec的表现优于目前最先进的顺序推荐模型。The existing sequence recommendation models have shortcomings in utilizing absolute interaction time,re‐sulting in inaccurate modeling of user preferences.Sequential recommendation based on contrast enhanced time-aware self-attention mechanism(CTiSASRec)was proposed.Firstly,the calculation process of attention weights inte‐grated rating data,absolute interaction time,location information,and project popularity.Secondly,the absolute inter‐action time and location order of the project were integrated to generate a new project location embedding.Finally,during the training process,contrast learning based on the results of two modeling sequences was used to distinguish the similarities and differences between samples,thereby improving the accuracy and robustness of the model.Experi‐mental studies conducted on six datasets of different fields and scales show that CTiSASRec outperforms state-of-the-art sequential recommendation models.

关 键 词：推荐系统 自注意力 时间感知模型 对比学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]