计算机应用 ›› 2017, Vol. 37 ›› Issue (5): 1392-1396.DOI: 10.11772/j.issn.1001-9081.2017.05.1392

• 人工智能 • 上一篇    下一篇

基于阶段时序效应的奇异值分解推荐模型

黄凯, 张曦煌   

  1. 江南大学 物联网工程学院, 江苏 无锡 214122
  • 收稿日期:2016-09-27 修回日期:2016-11-26 出版日期:2017-05-10 发布日期:2017-05-16
  • 通讯作者: 黄凯
  • 作者简介:黄凯(1988-),男,贵州毕节人,硕士研究生,CCF会员,主要研究方向:数据挖掘、推荐系统;张曦煌(1962-),男,江苏无锡人,教授,博士,主要研究方向:嵌入式系统、计算机网络。

Singular value decomposition recommender model based on phase sequential effect

HUANG Kai, ZHANG Xihuang   

  1. College of Internet of Things Engineering, Jiangnan University, Wuxi Jiangsu 214122, China
  • Received:2016-09-27 Revised:2016-11-26 Online:2017-05-10 Published:2017-05-16

摘要: 针对传统基于时序效应的奇异值分解(SVD)推荐模型在对用户预测评分建模过程中只考虑评分矩阵,采用复杂的时间函数拟合项目的生命周期、用户偏好的时序变化过程,造成模型难于解释、用户偏好捕获不准、评分预测精度不够高等问题,提出了一种改进的综合考虑评分矩阵、项目属性、用户评论标签和时序效应的推荐模型。首先,通过将时间轴划分时间段,利用sigmoid函数将项目的阶段流行度变换为[0,1]区间上的影响力来改进项目偏置;其次,利用非线性函数将用户偏置的时序变化转变为阶段评分均值与总体均值偏差的时序变化来改进用户偏置;最后,通过捕获用户对项目的阶段兴趣度,结合其相似用户在此时间段对该项目的好评率,生成用户项目交互作用影响因子,实现用户项目交互作用的改进。在Movielence 10M和20M电影评分数据集上的测试表明,改进模型能更好地捕获用户偏好的时序变化过程,提高评分预测准确性,均方根误差平均提高了2.5%。

关键词: 推荐系统, 时序效应, 奇异值分解, 项目流行度, 协同过滤

Abstract: The traditional Singular Value Decomposition (SVD) recommender model based on sequential effect only considers scoring matrix and uses complicated time function to fit item's life cycle and user's preferences, which leads to many problems, such as difficult to explain model, inaccurate to capture user's preferences and low prediction accuracy. In view of the drawbacks, an improved sequential effect model was proposed which considered scoring matrix, item attributes and user rating labels comprehensively. Firstly, the time axis was divided into different phases, the project's popularity was converted to influence in[0,1] to improve project bias by sigmoid function. Secondly, the time variation changes of the user bias were transformed into time variation changes of user rating mean and overall rating mean by nonlinear function. Finally, the influence factors of the user project interaction were generated to achieve the user project interaction improvement by capturing the user's interest, combining with favorable rate of the similar users. The tests on the Movielence 10M and 20M movie scoring data sets show that the improved model can better capture the time variation change of user preferences, improve the accuracy of scoring prediction, and improve the root mean square error by 2.5%.

Key words: recommender system, sequential effect, singular value decomposition(SVD), item popularity, collaborative filtering

中图分类号: