结合注意力机制与循环神经网络的序列推荐模型  

Sequence Recommendation Model Combining Attention Mechanism and Recurrent Neural Network

作　　者：吴晓 成卫青[1] WU Xiao;CHENG Weiqing(School of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing 210046,China)

机构地区：[1]南京邮电大学计算机学院,江苏南京210046

出　　处：《软件导刊》2023年第10期95-104,共10页Software Guide

基　　金：江苏省研究生教育教学改革项目(JGZZ19_038)。

摘　　要：在真实场景的推荐系统中,使用用户所有数据以训练推荐模型,存在忽略数据间时间顺序和学习到的用户兴趣是静态的问题。鉴于此,提出一种结合注意力机制与GRU循环神经网络的序列推荐模型,对用户的行为序列进行显式的动态建模,挖掘用户的长期偏好与短期偏好,并结合用户自身信息构建自适应分配权重的门控单元,将长期偏好与短期偏好结合,以此预测用户可能发生的下一个行为。在Amazon数据集上的实验结果表明,相较于目前基准推荐模型GRU4REC、STAMP、SASREC等,该模型在归一化折损累计增益(NDCG)和命中率(HIT)指标上分别至少提升了14.7%和8.8%,表明该模型能够更加准确地捕捉用户兴趣。In a real-world recommendation system,all user data is used to train the recommendation model,but there is a problem of ignor⁃ing the temporal order between data and the static learning of user interests.In view of this,a sequence recommendation model combining at⁃tention mechanism and GRU recurrent neural network is proposed,which explicitly and dynamically models the user´s behavior sequence,mines the user´s long-term and short-term preferences,and combines the user´s own information to construct an adaptive weighted gating unit to combine long-term and short-term preferences,in order to predict the next behavior that the user may experience.The experimental results on the Amazon dataset show that compared to the current benchmark recommendation models GRU4REC,STAMP,SASREC,etc.,the model has improved at least 14.7%and 8.8%in normalized discounted cumulative gain(NDCG)and hit rate(HIT)indicators,respectively,indicat⁃ing that the model can more accurately capture user interest.

关 键 词：序列推荐 注意力机制 循环神经网络 深度学习 长短期偏好 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]