【作者】 李鹏；

【导师】 于洪涛； 任传吉；

【作者基本信息】 燕山大学 ， 软件工程， 2013， 硕士

【摘要】 个性化协同过滤推荐系统作为最为成功的推荐系统之一，已经得到了许多电子商务公司的广泛应用。然而由于系统自身的开放性和用户的参与性，使它面临着安全性的严峻挑战。一些恶意用户出于商业利益的目的，向系统注入大量虚假信息，改变系统的推荐行为，从而降低了系统的推荐质量。因此，如何确保个性化协同推荐系统的安全性已成为亟待解决的问题。本文在对国内外研究现状综合分析的基础上，进一步对协同过滤推荐系统的安全问题进行了深入研究。首先，针对有监督的检测方法在用户概貌动态变化时检测模型不能进行增量式更新的问题，通过引入粗糙集理论提出一种基于增量学习的推荐攻击检测算法。该算法利用提出的基于边界样本的训练集生成算法选择最佳的用户标记样本建立训练集，以训练生成分类器。然后将每次采样中新加入的样本用于分类器的增量学习，在减少分类器训练时间的同时挖掘新样本的攻击特征，判断特征空间的分布区域，实现对攻击概貌样本更合理的覆盖。在此基础上，分类器利用基于统计特性的攻击检测方法检测出相应的攻击概貌。其次，针对现有的检测方法对真实概貌误判率较高的问题，提出一种基于多维风险因子的推荐攻击检测方法。该算法通过引入时间间隔可疑度、风险反馈、惩罚函数和用户评分活跃度等多个风险因子，从多个角度推理和评估用户评分行为的不确定性。然后，在多维风险因子的融合计算过程中，利用信息熵理论确立风险因子的分类权重，克服了过去常用的确定权重的主观判断方法。在此基础上，根据用户评分行为的风险评估值总和，检测出相应的攻击概貌。最后，对本文提出的两种攻击概貌检测算法进行了实验验证与分析，并与现有算法进行了对比研究，同时为今后的研究工作进行了展望。更多还原

【Abstract】 As one of the most successful recommendation system, collaborative filteringrecommendation system has a wide range of applications in e-commerce. Sincecollaborative recommender systems must be open to user input, it is difficult to design asystem that cannot be attacked. Some malicious users can provide a number of userprofiles to the system for some commerce purposes, trying to influence the system’sbehavior. This can make a serious impact on the recommendation quality of the system.So, how to solve the security problem of recommended system is urgently needed. On thebasis of the research and analysis about the present situation at home and abroad, thispaper has further conducted deep research of the security problem of recommendedsystem.Firstly, the existing supervised approaches suffer from the problem that detectionmodel cannot update incrementally with the increase of user profiles. Aiming at thisproblem, an algorithm for detecting recommendation attack based on incrementallearning is proposed. An algorithm for building a training set is proposed to choose thebest label samples used to build a classifier. Then newly labeled samples are used to trainthe classifier incrementally in order to make it more reasonable to cover attack profiles.Finally, the attack detection method based on statistical features is proposed todistinguish attack profiles from genuine user profiles.Secondly, according to the problem that the existing attack detection algorithmshave a high misjudgment rate for genuine user profiles, an algorithm for detectingrecommendation attack based on multiple risk factors is proposed, in which multiple riskfactors, including time suspicious degree, risk feedback, punish function, and activedegree, are incorporated to reflect uncertainty in various angles. Then the weight ofclassification is set up by information entropy theory for these risk factors, whichovercomes the shortage of traditional method, in which the weight is set up by subjectivemanners. Finally, the attack detection method distinguishes attack profiles from genuineuser profiles based on behaviors risk evaluation. Finally, we give the experimental evaluations and analysis of the algorithmsproposed in this paper, compare the performance between the proposed algorithms andother existing algorithms, and make the conclusions and prospects for the further search.更多还原