融合时间上下文信息的序列推荐系统框架  

Sequence Recommender System Framework Incorporating Temporal Context Information

作　　者：逯暄 昝晓亮 彭甫镕 颜无瑕 LU Xuan;ZAN Xiaoliang;PENG Furong;YAN Wuxia(College of Physical and Electronic Engineering,Shanxi University,Taiyuan 030006,China;Institute of Big Data Science and Industry,Shanxi University,Taiyuan 030006,China;College of Mathematics and Information Science,Nanjing Normal University of Special Education,Nanjing 210038,China)

机构地区：[1]山西大学物理电子工程学院,太原030006 [2]山西大学大数据科学与产业研究院,太原030006 [3]南京特殊教育师范学院数学与信息科学学院,南京210038

出　　处：《小型微型计算机系统》2025年第4期796-802,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金面上项目(62276162)资助;山西省基础研究计划项目(202203021222016)资助;江苏省高等学校自然科学研究面上项目(19KJB520011)资助.

摘　　要：序列推荐相对于协同过滤方法在考虑用户行为顺序和个性化推荐等方面具有明显优势.然而现有的大多数序列推荐模型着重关注用户历史行为间的关系,忽略了用户兴趣随时间的演化特征,对此,本文提出了一种融合时间上下文信息的序列推荐通用学习框架.该框架将Tucker分解技术与注意力机制相融合,为每个用户行为分配适当的注意力权重,使模型关注对当前推荐任务更重要的时间上下文信息.此外,使用了分布鲁棒性损失函数来解决训练数据与测试数据之间可能存在的分布漂移问题.3个公开数据集上的实验结果表明,本文框架适用于不同的序列推荐模型并能提升推荐性能.Compared with collaborative filtering methods,sequence recommendation has obvious advantages in considering the order of user behavior and personalized recommendation.However,most of the existing sequence recommendation models focus on the relationship between users′historical behaviors and ignore the evolution of users′interests over time.In this paper,a general learning framework for sequence recommendation that fuses temporal context information is proposed.The framework integrates the Tucker decomposition technique with the attention mechanism,assigns appropriate attention weights to each user behavior,enabling the model to focus on temporal context information that is more important for the current recommendation task.In addition,The distribution robustness loss function is used to solve the possible distribution drift between the training data and the test data.Experimental results on three public datasets show that the proposed framework is suitable for different sequence recommendation models and can improve the recommendation performance.

关 键 词：推荐系统 序列推荐 时间上下文信息 Tucker分解 分布漂移 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]