基于可疑用户度量的鲁棒推荐方法研究

发布时间：2018-04-23 10:11

本文选题：协同过滤 + 鲁棒推荐　；参考：《燕山大学》2016年博士论文

【摘要】：协同过滤推荐系统被广泛地应用到电子商务网站等诸多领域,可以有效解决“信息超载”问题。但是,一些恶意用户蓄意伪造虚假用户评分来干扰系统的决策推荐过程,企图使系统产生有利于个人的推荐结果,这种恶意攻击行为严重影响了系统的推荐质量以及用户对系统的信任。因此,如何保障推荐系统不受恶意攻击的影响,为用户提供真实可靠的推荐结果已经成为一个值得研究的热点问题。本文基于可疑用户度量的思想,从基于内存和基于模型的推荐技术两方面展开研究,致力于设计一系列鲁棒性高、精度损失少的协同过滤推荐算法。首先,针对基于用户的推荐算法近邻选取可靠性不高的问题,提出一种基于k-距离和项目类别信息的鲁棒推荐方法。根据离群点检测思想,实现用户可疑度计算;将用户可疑度与项目类别信息相融合,给出缺失值填充计算方法,对用户的未评分项进行填充;基于填充后的评分矩阵,结合传统的基于用户的协同过滤推荐技术将用户相似度和可疑度共同作为选取邻居的依据,实现对目标用户的鲁棒推荐。其次,针对已有信任计算模型在攻击概貌存在情况下对用户间信任关系度量不准确的问题,提出一种基于可疑用户度量和多维信任的鲁棒推荐方法。根据用户概貌的特征训练相关向量机分类器,对用户可疑度进行度量;基于用户评分信息挖掘用户之间的隐式信任关系,结合用户可疑性信息构建可靠多维信任模型;将可靠多维信任模型与基于用户的近邻推荐模型相融合,完成对目标用户的可靠推荐。再次,针对基于矩阵分解的推荐算法在面对托攻击时鲁棒性较差的问题,提出一种基于模糊核聚类和支持向量机的鲁棒推荐方法。根据攻击概貌间高相似度的特性,利用模糊核聚类技术在高维特征空间对用户概貌进行聚类,将攻击概貌聚到同一类内;利用支持向量机分类器对含有攻击概貌的聚类进行检测,进一步识别攻击概貌;将攻击概貌识别结果融入到矩阵分解过程中,提高算法的鲁棒性。然后,针对基于矩阵分解的推荐算法不能平衡处理鲁棒性和推荐精度的问题,提出一种基于可疑用户识别和Tukey M-估计量的鲁棒推荐方法。根据用户评分信息的分布情况,提出评分个数偏离度和邻居平均相似度的计算方法,对可疑用户进行识别,将识别结果与传统的近邻选取思想相结合,构建可靠近邻模型;在矩阵分解过程中引入Tukey M-估计量,构造鲁棒矩阵分解模型;将可靠近邻模型融入到鲁棒矩阵分解模型中,在提高算法鲁棒性的同时提高推荐精度。最后,在MovieLens数据集上与现有的经典方法进行了实验对比分析,验证了所提方法的有效性。
[Abstract]:Collaborative filtering recommendation system is widely used in many fields, such as e-commerce websites, which can effectively solve the problem of "information overload". However, some malicious users deliberately falsify false user ratings to interfere with the decision-making and recommendation process of the system, in an attempt to make the system produce recommendations in the interests of individuals. This malicious attack seriously affects the recommendation quality of the system and user's trust in the system. Therefore, how to protect the recommendation system from malicious attacks and provide users with reliable recommendation results has become a hot issue worthy of study. Based on the idea of suspect user metrics, this paper studies the memory and model-based recommendation techniques, and designs a series of collaborative filtering recommendation algorithms with high robustness and low precision loss. Firstly, a robust recommendation method based on k- distance and item category information is proposed to solve the problem of low reliability of nearest neighbor selection based on user-based recommendation algorithm. According to the idea of outlier detection, the user suspect degree can be calculated; the missing value filling calculation method is given by combining the user suspicious degree with item category information; based on the filled score matrix, Combined with the traditional user-based collaborative filtering recommendation technology, the similarity and suspicious degree of users are taken as the basis for selecting neighbors, and the robust recommendation to target users is realized. Secondly, a robust recommendation method based on suspect user metrics and multidimensional trust is proposed to solve the problem of inaccurate measurement of trust relationships between users in the presence of existing trust computing models. According to the features of the user profile, the correlation vector machine classifier is trained to measure the degree of user suspicion, the implicit trust relationship between users is mined based on the user score information, and the reliable multi-dimensional trust model is constructed by combining the user suspicious information. The reliable multi-dimension trust model is combined with the user-based nearest neighbor recommendation model to complete the reliable recommendation to the target user. Thirdly a robust recommendation method based on fuzzy kernel clustering and support vector machine is proposed to solve the problem of poor robustness of the recommendation algorithm based on matrix decomposition. According to the characteristics of high similarity between attack profiles, fuzzy kernel clustering technology is used to cluster the user profile in high dimensional feature space, and the attack profile is clustered into the same class. Support vector machine (SVM) classifier is used to detect the cluster with attack profile, to further identify the attack profile, and to incorporate the result of attack profile recognition into matrix decomposition process, so as to improve the robustness of the algorithm. Then, a robust recommendation method based on suspect user identification and Tukey M- estimator is proposed to solve the problem that the recommendation algorithm based on matrix decomposition can not deal with the problem of robustness and recommendation accuracy. According to the distribution of the users' rating information, the method of calculating the number deviation of the score and the average similarity of the neighbors is put forward. The suspicious users are identified, and the identification results are combined with the traditional idea of nearest neighbor selection to construct the reliable nearest neighbor model. In the process of matrix decomposition, Tukey M- estimator is introduced to construct the robust matrix decomposition model, and the reliable nearest neighbor model is incorporated into the robust matrix decomposition model, which improves the robustness of the algorithm and improves the recommendation accuracy. Finally, the effectiveness of the proposed method is verified by comparing with the existing classical methods on the MovieLens dataset.
【学位授予单位】：燕山大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】