计算机工程 ›› 2019, Vol. 45 ›› Issue (8): 198-202,209.doi: 10.19678/j.issn.1000-3428.0051294

• 人工智能及识别技术 • 上一篇    下一篇

基于循环神经网络的推荐算法

高茂庭, 徐彬源   

  1. 上海海事大学 信息工程学院, 上海 201306
  • 收稿日期:2018-04-23 修回日期:2018-07-23 出版日期:2019-08-15 发布日期:2019-08-08
  • 作者简介:高茂庭(1963-),男,教授、博士,主研方向为智能信息处理、数据库与信息系统;徐彬源,硕士研究生。
  • 基金项目:
    国家自然科学基金(61202022)。

Recommendation Algorithm Based on Recurrent Neural Network

GAO Maoting, XU Binyuan   

  1. Collage of Information Engineering, Shanghai Maritime University, Shanghai 201306, China
  • Received:2018-04-23 Revised:2018-07-23 Online:2019-08-15 Published:2019-08-08

摘要: 传统电影推荐算法多数基于用户和电影的静态属性进行推荐,忽略了时间序列数据内在的时间和因果因素,推荐质量不高。为此,利用循环神经网络(RNN)在处理时间序列上的优势,提出一种推荐算法R-RNN。采用2个长短期记忆网络分别挖掘用户和电影的潜在状态,实现长距离的历史状态积累,将用户状态和电影状态的内积作为最终评分。在IMDB和Netflix数据集及Netflix子集上的实验结果表明,与基于概率矩阵分解、TimeSVD++及AutoRec算法相比,该算法能够有效降低均方根误差,并提高预测评分的准确度。

关键词: 推荐算法, 循环神经网络, 长短期记忆网络, 时间动态, 潜在状态

Abstract: The traditional movie recommendation algorithms mostly make recommendations based on the static attributes of users and movies,which ignores the inherent temporal and causal relationship in the time series data.The recommendation quality need to be improved.To this end,taking the advantage of Recurrent Neural Network (RNN) in processing time series,this paper proposes a recommendation algorithm R-RNN.Two Long-Short Term Memorys (LSTMs) are used to mine the potential status of users and films,long historical status accumulation is realized,and the prediction score is obtained by the inner product of the user status and the movie status.Experiments on IMDB and Netflix datasets and Netflix subsets show that compared with those of the probability matrix decomposition,TimeSVD++ and AutoRec algorithms,the algorithm can effectively reduce the root mean square error and improve the accuracy of the rating prediction.

Key words: recommendation algorithm, Recurrent Neural Network(RNN), Long-Short Term Memory(LSTM), time dynamics, potential status

中图分类号: