【作者】 张富国；

【导师】 徐升华；

【作者基本信息】 江西财经大学 ， 管理科学与工程， 2009， 博士

【摘要】 随着互联网的蓬勃发展,电子商务的虚拟购物环境既为企业提供了新的发展机遇,也给用户提出了如何处理Web商品信息过载问题的挑战,推荐系统作为个性化服务的一种方式,能够向用户推荐其感兴趣的项目,辅助用户作出决策,成为用户网上购物的有力助手。自20世纪90年代首次出现推荐系统以来,对此领域的研究虽然也取得了非常大的进步,但传统的协同过滤推荐技术依然存在着数据稀疏性、冷启动、“托”攻击、黑匣子和伸缩性等多个难以克服的问题,而在Web社会网络迅速发展的形势下,基于信任的电子商务推荐系统由于在传统的协同过滤推荐系统中引入了信任机制,能有效改善或克服以上缺陷,成为目前推荐系统研究最重要的课题之一。围绕着如何提高用户对推荐系统的推荐满意度,本文对电子商务环境下基于信任的个性化推荐系统的若干关键问题进行了有益的探索和研究。论文首先系统地综述了目前国内外在个性化推荐领域的理论研究和发展现状,重点分析协同过滤推荐方法的优势和存在的研究热点问题,继而从网络信任的定义和特性出发,对当前多种信任度评估模型的特点进行了深刻分析,在此基础上,作者提出了基于信任的个性化推荐系统的一般模型和形式化表示方法,探讨了新系统的框架结构;接着本文以如何提高用户对推荐系统的推荐满意度为主线,分别对用户多兴趣下基于信任的个性化推荐算法、推荐列表的多样性和推荐系统的推荐攻击问题进行了系统的研究。本论文的研究创新主要体现在以下几个方面:(1)借鉴传统协同过滤系统的基础模型以及Web社会网络的特点,提出了基于信任的电子商务个性化推荐系统的一般模型,该模型具有很好的包容性,可以扩展出不同的推荐算法。(2)在分析了目前概貌级信任模型不适合用户多兴趣情况下项目推荐的基础上,提出了基于主题级信任模型的协同过滤推荐算法,并通过实验验证,新算法能有效解决多兴趣问题。(3)提出了基于信任的推荐多样性算法,该算法通过选择主题多样的信任邻居来平衡推荐结果的准确性和多样性。并通过一系列的实验结果表明,该算法能有效地提高推荐的多样性。(4)在对信任传递可能引起的恶意攻击形式进行分析的基础上,提出了用数据起源法来追踪恶意用户,进而把恶意用户限制为不信任用户的防御方法。更多还原

【Abstract】 With the rapid popularization of Internet, virtual shopping environment of e-commerce not only provide new development chance for the enterprises, but also put forward how to handle the challenge of information overload of web products. As a means of personalized service, recommender systems can recommend interesting items for users, assist users to make decision, and become an good assistant for users’ shopping online. Since the emergence of the first recommender system in the 1990s, remarkable progress in this field has been achieved, but there still exists some issues hard to overcome in traditional collaborative filtering technology. In the situation of the rapid development of web social networks, because of incorporating trust mechanism into traditional collaborative filtering recommender systems, trust-based recommender systems can improve or overcome these limitations effectively, and become one of the most important research projects.This thesis explored and researched some key issues of trust-based personalized recommender systems in e-Commerce. Firstly, it gave an overview of the theoretical research and development on personalized recommendation to date. The advantages of collaborative filtering technology and research challenges were analyzed. Secondly, from the definition and properties of network trust, the characteristic of the trust evaluation models were analyzed and compared. Thirdly, a general model of trust-based personalized recommender system and system framework was presented. Then the author gave the research work on trust-based personalized recommendation algorithm for user’s multiple interests, diversity of recommendation list items and recommendation attack of trust-based recommender system.The innovational works of this thesis are as following:(1) Based on the basic modal of traditional collaborative filtering system and the characteristic of web social network, the author presented a general modal of trust-based e-commerce personalized recommender systems, which is inclusive, and can extend to many different recommendation algorithm. Meanwhile, the system framework of trust-based e-commerce personalized recommender system is discussed.(2) Finding that current profile-level trust modal is not suitable to items recommendation for user’s multiple-interests, the authors analyzed the reason and present a novel collaborative filtering algorithm based on topic-level trust modal, which can deal with the issue of user’s multiple-interests. A series of experiments show that the algorithm achieves highly on the accuracy and robustness under shilling attacks.(3) Considering the user’s complete spectrum of interests, the limitation of research on recommender system only paying attention to improve the accuracy of recommendation algorithm is the neglect of recommendation diversification. This paper proposes a novel recommendation diversification algorithm for trust based E-commerce personalized recommender systems, which is designed to balance the accuracy and the diversification of the recommendation list. The experiment shows that the algorithm can improve the recommendation diversification.(4) Malicious attack modes because of trust propagation was discussed. According to the sparsity of rating matrix of trust neighors, the author presented a defending method by tracing malicious users with data lineage, and limiting the malicious users to distrust users.更多还原