基于非线性特征和Cauchy加权M-估计量的鲁棒推荐算法

发布时间：2018-06-25 19:55

本文选题：托攻击 + 鲁棒推荐算法　；参考：《计算机学报》2017年06期

【摘要】：协同推荐系统作为一种重要的个性化服务模式,在电子商务站点中的应用越来越广泛.然而,各种恶意欺骗和虚假反馈已制约了其应有效能的发挥.如何有效识别和抵御托攻击,确保系统推荐的可信性,是推荐系统研究面临的一大挑战.为了确保推荐的可信性,人们提出了各种基于矩阵分解的鲁棒协同推荐算法.但是这些推荐算法在面对托攻击时不仅精度损失大,而且鲁棒性较差.为此,该文提出一种基于非线性特征和Cauchy加权M-估计量的鲁棒推荐算法.首先,采用核主成分分析方法提取用户评分矩阵的非线性特征,以充分挖掘推荐系统中用户(或项目)之间的内在关联,最大限度地保留用户和项目的特征信息,提高推荐精度和鲁棒性;然后,引入鲁棒统计中的Cauchy加权M-估计量,并联合矩阵分解模型对用户和项目特征矩阵进行鲁棒参数估计,以限制攻击概貌对参数估计过程产生的影响;最后,设计相应的鲁棒协同推荐算法,并在MovieLens和Netflix数据集上对算法的有效性进行评价.实验结果表明:该文算法在推荐精度和鲁棒性方面明显优于现有的鲁棒推荐算法.
[Abstract]:As an important personalized service mode, collaborative recommendation system is more and more widely used in e-commerce sites. However, all kinds of malicious deception and false feedback have restricted its effectiveness. How to effectively identify and resist trust attacks and ensure the credibility of the system recommendation system is a major challenge to the research of recommendation system. In order to ensure the credibility of recommendation, various robust collaborative recommendation algorithms based on matrix decomposition are proposed. However, these recommended algorithms not only have a large loss of precision but also have poor robustness in the face of support attacks. In this paper, a robust recommendation algorithm based on nonlinear features and Cauchy weighted M- estimator is proposed. Firstly, the kernel principal component analysis (KPCA) method is used to extract the nonlinear features of the user scoring matrix in order to fully mine the internal relationship between the users (or items) in the recommendation system, and to retain the feature information of the user and the item to the maximum extent. Then, the Cauchy weighted M- estimator in robust statistics is introduced, and the robust parameter estimation of user and item characteristic matrix is carried out by combining matrix decomposition model. Finally, a robust collaborative recommendation algorithm is designed and evaluated on Movie Lens and Netflix datasets. Experimental results show that the proposed algorithm is superior to the existing robust recommendation algorithm in recommendation accuracy and robustness.
【作者单位】：燕山大学信息科学与工程学院;河北省计算机虚拟技术与系统集成重点实验室(燕山大学);
【基金】：国家自然科学基金(61379116) 河北省自然科学基金(F2013203124,F2015203046) 河北省高等学校科学技术研究重点项目基金(ZH2012028)资助~~
【分类号】：TP391.3

【相似文献】