计算机应用 ›› 2015, Vol. 35 ›› Issue (12): 3497-3501.DOI: 10.11772/j.issn.1001-9081.2015.12.3497

• 人工智能 • 上一篇    下一篇

基于自适应提升的概率矩阵分解算法

彭行雄1,2, 肖如良1,2, 张桂刚3   

  1. 1. 福建师范大学软件学院, 福州 350117;
    2. 大数据分析与应用福建省高校工程研究中心, 福州 350117;
    3. 中国科学院自动化研究所, 北京 100190
  • 收稿日期:2015-06-02 修回日期:2015-08-25 出版日期:2015-12-10 发布日期:2015-12-10
  • 通讯作者: 肖如良(1966-),男,湖南娄底人,教授,博士,主要研究方向:大数据云服务
  • 作者简介:彭行雄(1991-),男,湖北孝感人,硕士研究生,主要研究方向:机器学习;张桂刚(1978-),男,湖南邵阳人,副研究员,博士,主要研究方向:云计算、海量数据处理。
  • 基金资助:
    教育部规划基金项目(11YJA860028);福建省科技计划重大项目(2011H6006)。

Probabilistic matrix factorization algorithm based on AdaBoost

PENG Xingxiong1,2, XIAO Ruliang1,2, ZHANG Guigang3   

  1. 1. Faculty of Software, Fujian Normal University, Fuzhou Fujian 350117, China;
    2. Fujian Provincial University Engineering Research Center of Big Data Analysis and Application, Fuzhou Fujian 350117, China;
    3. Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China
  • Received:2015-06-02 Revised:2015-08-25 Online:2015-12-10 Published:2015-12-10

摘要: 针对推荐系统中概率矩阵分解模型(PMF)泛化能力(对新用户和物品的推荐性能)较差、预测准确性不高的问题,提出一种新的基于自适应提升的概率矩阵分解算法(AdaBoostPMF)。该算法首先为每个样本分配样本权重;然后根据PMF中的每一轮随机梯度下降法学习用户和物品特征向量,并计算总体预测误差均值和标准差。从全局的角度利用AdaBoost思想自适应调整样本权重,使算法更注重学习预测误差较大的样本;最后对预测误差分配样本权重,让用户和物品特征向量找到更合适的优化方向。相比传统的PMF算法,AdaBoostPMF算法能够将预测精度平均提高约2.5%。实验结果表明,该算法通过加权预测误差较大的样本,能够较好地拟合用户特征向量和物品特征向量,提高预测精度,可以有效地应用于研究个性化推荐。

关键词: 推荐系统, 概率矩阵分解, 自适应提升, 模型融合, 评分预测

Abstract: Concerning the poor generalization ability (the recommended performance for new users and items) and low predictive accuracy of Probabilistic Matrix Factorization (PMF) in recommendation system, a new algorithm of Probabilistic Matrix Factorization algorithm based on AdaBoost (AdaBoostPMF) was proposed. Firstly, the initial weight for each sample was assigned. Secondly, the feature vectors of users and items were learned by each round of PMF stochastic gradient descent method and the global mean and standard deviation of the prediction error were calculated. The sample weights were adaptively adjusted by using AdaBoost from the a global perspective, which made the proposed algorithm pay more attention to training those samples with the larger prediction error than others. Finally, the sample weights were assigned to predictive error, which found the more appropriate optimum direction for feature vectors of users and items. Compared with traditional PMF algorithm, the proposed AdaBoostPMF algorithm could significantly improve the prediction precision by about 2.5% on average. The experimental results show that, the proposed algorithm can better fit the user feature vector and the item feature vector and improve the prediction accuracy by weighting the samples with larger prediction error.The proposed algorithm can be effectively applied to the personalized recommendation.

Key words: recommendation system, Probabilistic Matrix Factorization (PMF), AdaBoost, model blending, rating prediction

中图分类号: