融合用户兴趣建模的智能推荐算法研究  

Research on intelligent recommendation algorithm integrating user interest modeling

作　　者：洪志理 赖俊 曹雷 陈希亮 Hong Zhili;Lai Jun;Cao Lei;Chen Xiliang(Command&Control Engineering College,Army Engineering University of PLA,Nanjing 210007,China)

机构地区：[1]陆军工程大学指挥控制工程学院,江苏南京210007

出　　处：《信息技术与网络安全》2021年第11期37-48,共12页Information Technology and Network Security

摘　　要：强化学习被越来越多地应用到推荐系统中。提出一种基于DDPG融合用户动态兴趣建模的推荐方法(DDPG-LA),使用LSTM网络提取用户的长期兴趣,利用注意力机制方法提取用户的短期兴趣,将两种兴趣结合作为智能体的状态。同时,在LSTM网络中加入状态增强单元,以加速模型对于用户长期兴趣的建模,在注意力机制中加入缓解推荐延迟的模块来解决该方法应用于推荐系统中时所产生的缺陷。在Movelines的两个数据集上对模型进行实验,同时在各种测试指标上与传统方法进行比较,结果显示所提出的算法更具优越性。Reinforcement learning is more and more applied to recommendation system.This paper proposes a recommendation method based on DDPG and user dynamic interest modeling(DDPG-LA).It uses LSTM network to extract user′s long-term interest and attention mechanism to extract user′s short-term interest.The two kinds of interest are combined as the state of agent.At the same time,the state enhancement unit is added to LSTM network to accelerate the modeling of users′long-term interest,and the module to alleviate the recommendation delay is added to the attention mechanism to solve the defects when the method is applied to the recommendation system.In this paper,the model is tested on two data sets of Movelines,and compared with the traditional methods in various test indexes,the results show that the proposed algorithm has more advantages.

关 键 词：强化学习 推荐系统 DDPG DDPG-LA LSTM 注意力机制 长期兴趣 短期兴趣 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]