计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (1): 175-188.DOI: 10.3778/j.issn.1673-9418.2211022

• 人工智能·模式识别 • 上一篇    下一篇

时间感知的双塔型自注意力序列推荐模型

余文婷,吴云   

  1. 1. 贵州大学 公共大数据国家重点实验室,贵阳 550025
    2. 贵州大学 计算机科学与技术学院,贵阳 550025
  • 出版日期:2024-01-01 发布日期:2024-01-01

Time-Aware Sequential Recommendation Model Based on Dual-Tower Self-Attention

YU Wenting, WU Yun   

  1. 1. The State Key Laboratory of Public Big Data, Guizhou University, Guiyang 550025, China
    2. School of Computer Science and Technology, Guizhou University, Guiyang 550025, China
  • Online:2024-01-01 Published:2024-01-01

摘要: 用户的偏好具有聚合性和漂移性。现有推荐算法在序列建模框架中融合了交互时间相关性的建模,取得了很大的性能改善,但它们在建模时仅考虑了交互的时间间隔,使得它们在捕捉用户偏好的时间动态方面存在局限性。首先,提出了一种新的时间感知的位置嵌入方法,将时间信息与位置嵌入相结合,帮助模型学习时间层面的项目相关性。随后,在时间感知位置嵌入基础上,提出了时间感知的双塔自注意力序列推荐模型(TiDSA)。TiDSA包含项目级和特征级的自注意力模块,分别从项目和特征两个角度对用户偏好随时间变化的过程进行分析,实现了对时间、项目和特征的统一建模,并且在特征级自注意力模块,设计了多维度的自注意力权重计算方式,从特征维度、项目维度和项目与特征交叉维度充分学习特征之间的相关性。最后,TiDSA将项目级与特征级的信息相融合得到最终的用户偏好表示,并根据该表示为用户提供可靠的推荐结果。四个真实推荐数据集的实验结果表明,TiDSA的性能优于许多先进的基线模型。

关键词: 时间感知序列推荐, 位置嵌入, 特征级自注意力机制, 双塔自注意力网络

Abstract: Users’ preferences are migratory and aggregated. Although recommenders have been greatly improved by modeling the timestamps of interactions within a sequential modeling framework, they only consider the time interval of interactions when modeling, making them limited in capturing the temporal dynamics of user prefer-ences. For this reason, this paper proposes a novel time-aware positional embedding that fuses temporal information into the positional embedding to help the network learn item correlations at the temporal level. Then, based on the time-aware positional embedding, this paper proposes a time-aware sequential recommendation model based on dual-tower self-attention (TiDSA). TiDSA includes item-level and feature level self-attention blocks, which analyzes the process of user preference change over time from the perspective of items and features respectively, and achieves the unified modeling of time, items and features. In addition, in the feature-level self-attention block, this paper calculates the self-attention weights from three dimensions, namely, feature-feature, item-item and item-feature, to fully capture the correlation between different features. Finally, the model fuses the item-level and feature-level information to obtain the final user preference representation and provides reliable recommendation results for users. Experimental results on four real-world datasets show that TiDSA outperforms various state-of-the-art models.

Key words: time-aware sequential recommendation, positional embedding, feature-level self-attention, dual-tower self-attention