基于增量式聚类和矩阵分解的鲁棒推荐方法  

Robust Recommendation Approach Based on Incremental Clustering and Matrix Factorization

作　　者：徐玉辰[1,2] 刘真[1,2] 张付志[1,2] 

机构地区：[1]燕山大学信息科学与工程学院,秦皇岛066004 [2]河北省计算机虚拟技术与系统集成重点实验室,秦皇岛066004

出　　处：《小型微型计算机系统》2015年第4期689-695,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61379116)资助;河北省自然科学基金项目(F2011203219;F2013203124;F2015203046)资助;河北省高等学校科学技术研究重点项目(ZH2012028)资助

摘　　要：现有基于矩阵分解的推荐算法在面对"托"攻击时鲁棒性较差,而且随着用户和项目的增加需要进行模型重构,时间代价较高.针对上述问题,提出一种基于增量式聚类和矩阵分解的鲁棒推荐方法.首先,采用基于k-median聚类的用户评分矩阵分块算法对用户评分矩阵中的相似用户进行聚类,构建用户评分矩阵分块,并对每个分块进行鲁棒矩阵分解建立推荐模型;然后,针对新增用户(项目),利用增量式聚类算法对用户评分矩阵分块进行更新,并采用基于加权信息熵的线性回归方法,对新增用户(项目)的特征向量进行局部参数估计.在Movie Lens 1M数据集上的实验结果表明,本文提出的推荐方法不仅具有较高的推荐精度和鲁棒性,而且模型更新的时间明显减少.The existing recommendation algorithms based on matrix factorization expose low robustness when facing the shilling at- tacks. Moreover, with the increase of users and items, the existing recommendation models need to be reconstructed, which is always time consuming. Aiming at the problems above, we propose a robust recommendation approach based on incremental clustering and matrix factorization. Firstly,we use user rating matrix block algorithm based on k-median clustering to cluster the similar users and construct the user rating matrix blocks. Then we use matrix factorization for each block to construct the recommendation model. Sec- ondly, for the new users or items, we use incremental clustering algorithm to update the user rating matrix blocks and adopt the linear regression approach based on weighted entropy to realize the local parameter estimation. Experimental results on the MovieLens 1 M dataset show that the proposed approach exhibits the high recommendation accuracy and robustness, moreover the time of model update is reduced obviously.

关 键 词：鲁棒推荐算法 增量式聚类 矩阵分解 信息熵 模型更新 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]