融合隐性特征的群体推荐方法研究  

Research on Method of Group Recommendation for Fusion of Hidden Features

作　　者：刘毅[1] 钟忺[1] 李琳[1] LIU Yi ZHONG Xian LI Lin(Department of Computer Science and Technology,Wuhan University of Teehnology,Wuhan 430070, Chin)

机构地区：[1]武汉理工大学计算机科学与技术学院,武汉430070

出　　处：《计算机科学》2017年第3期231-236,共6页Computer Science

基　　金：国家社会科学基金(15BGL048);武汉市创新团队项目(201307020402005);湖北省自然科学基金:城市隧道监控大数据分析预测及视频语义检索方法研究(ZRY2015001126)资助

摘　　要：作为目前最成功的主流推荐方法,奇异值分解算法(SVD)将已知的海量数据建模并通过矩阵分解降维处理来得到有效信息;非负矩阵分解(NMF)则通过分解出非负矩阵元素来解释特征意义。这两种较为成功的方法均通过对显性反馈信息进行基于矩阵分解的处理得到用户的喜好信息来进行群体推荐。然而,仅凭用户的显性反馈信息有时无法准确反映用户的真实喜好。为解决上述问题,提出了一种针对这两种模型的改进方法,将隐性特征和基于隐性特征的群体权重计算方法融合进经典的矩阵分解算法,其中隐性特征可以完善用户的喜好信息,基于隐性特征的群体权重计算方法则根据群体的特点给予用户相应的权重,使得推荐的准确率得到提升。对该方法在KDD Cup 2012Track1中的腾讯微博数据集上进行测试,实验结果表明在该数据集上融合方法的平均绝对偏差(MAE)和准确率要优于SVD算法与NMF算法,推荐的性能有较明显的提升。As the most successful mainstream recommendation method, singular value decomposition （SVD） algorithm builds the model from known huge data and uses the matrix decomposition dimension reduction to get effective informa- tion, and non negative matrix factorization （NMF） uses the decomposition of nonnegative matrix elements to explain the meanings of characteristics. These two kinds of successful methods are based on matrix decomposition of explicit feed- back information, and obtain the user~ s preference information. However, they cannot accurately reflect the true preferences of the users only according to user＇ s explicit feedback. To solve the problem, this paper put forward an im- proved method for the two models, integrating the hidden features and weight calculation method based on hidden fea- tures into the classical matrix decomposition algorithm, the hidden features can perfect the information of user＇s prefe- rences. Weight calculation method based on hidden features can judge the group characteristics and give the appropriate weight to users, which improve the recommendation accuracy. The method was tested on the Tencent micro blog data set in KDD Cup 2012 Trackl. The results show that from the experimental standard of the MAE and the precision, the fusion method is better than SVD and NMF on this data set, and significantly improves the recommendation perfor- mance.

关 键 词：群体推荐 隐性特征 群体权重 平均绝对偏差 

分 类 号：TP319.41[自动化与计算机技术—计算机软件与理论]