【作者】 张宇；

【导师】 吴朝晖； 陈华钧；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2009， 博士

【摘要】 随着信息技术的飞速发展和互联网的广泛普及,电子邮件、在线购物、在线交友、网上支付、即时通讯等应用已经成为人们工作和生活中不可或缺的一部分。人们在互联网上通过各种关系相互联系在一起,形成了一个个规模庞大、关系复杂以及内容丰富的在线社会网络。面向商业领域的实际应用,目前对在线社会网络的研究面临两大挑战。一是,由于互联网的开放性和匿名性,如何给在线社会网络中的用户提供合理的信任计算机制成为一个亟待解决的问题。二是,在线社会网络作为一个新的商业渠道,面向实际应用,如何对在线社会网络进行挖掘分析从而取得更好的经济效益和社会效益也成为一个备受关注的焦点问题。这两方面直接关系到在线社会网络的安全性和实用性,因此对其研究既具有理论价值又具有实际意义。本文正是以在线社会网络的迅猛发展为背景,针对上述两方面开展深入研究,提出了相应的模型和算法。本文的研究内容和主要贡献有以下几个方面:·基于在线社会网络的链式信任模型根据用户可以在社会网络中传递信任消息的特点,本文提出了一个集成信任值和置信值的二维链式信任模型。面向不同的应用场景和用户个性化需求,给出了多种信任计算策略。针对互联网的开放性和匿名性,介绍了链式信任模型的局部存储机制、动态更新机制和信任报告机制。通过在Epinions网站的真实在线社会网络上进行实验,证明了链式信任模型的有效性和灵活性。·基于语义的信任推理本文创新性地提出了一种基于语义的信任推理机制。该机制利用语义网技术定义了Epinions领域本体,然后采用OWL/RDF语言对Epinions网站上的数据进行知识表达,并从中抽取出信任相关的信息。根据这些信息,本文制定了基于OWL的信任推理规则。利用这些规则,我们可以推理出用户感兴趣的领域,并从泛化的信任关系推理出领域相关的信任关系,以及根据用户的反馈行为推理出隐式信任关系,从而支持更加准确高效的信任计算。·集成声望、内容和上下文信息的组合信任模型绝大多数现有的信任模型都是基于声望信息的,但是仅仅利用声望信息进行信任计算是远远不够的。因此,本文提出了一个集成声望、内容和上下文信息的组合信任模型——RCCtrust。RCCtrust利用基于语义的信任推理机制从互联网的内容和上下文中抽取出信任相关的信息,并将用户在商品评分和反馈行为上的相似度集成起来刻画用户之间的信任程度,从而构建出一个边权重的组合信任网络来进行信任计算。实验结果表明RCCtrust模型无论是在准确率方面,还是在覆盖率方面,都优于传统协同过滤的单纯相似度方法和仅仅利用信任关系的trust-aware方法。·基于启发式信息的目标团体发现本文提出了一个在社会网络中基于启发式信息进行目标团体发现的算法。根据对Epinions网站上的在线社会网络进行分析得出的特征,并结合用户的角色和在线行为,抽取出进行目标团体发现的启发式信息,从而可以有效地简化算法和减小搜索空间。实验结果显示,基于启发式信息的目标团体发现算法能够有效地发现在线社会网络中具有影响力的目标团体,在基于社会网络的市场营销中具有广泛的应用价值。·基于影响力最大化的关键人物挖掘传统关键人物挖掘算法往往只考虑了社会网络的结构特点,而忽略了节点之间的相互作用。针对这一不足,本文提出了一个基于影响力最大化的关键人物挖掘算法。面向Epinions网站上的信任网络,利用从互联网的内容和上下文中挖掘出的信息,对用户之间的影响力关系进行建模,并通过求解影响力最大化问题来确定在线社会网络中的关键人物。实验结果表明,本文提出的基于影响力最大化的爬山算法在不同的阈值区间都优于其它几种算法,特别是在激活阈值较大的情况下,爬山算法的优势更为明显。更多还原

【Abstract】 With the rapid development of information technology and widespread use of Internet, applications such as email,online shopping,online social networking,online payment and instant messaging have become an indispensable part of people’s work and life.People ple are connected to each other through a variety of mutual relationships,forming many large,complicated and content-rich online social networks.For practical applications in business,current research of online social networks are facing two major challenges.First, due to the openness and anonymity of the lnternet,how to provide reasonable trust computing mechanism for users in online social networks is an urgent problem to solve.Second, online social network serving as a new business channel,how to conduct mining analysis on online social networks to achieve better economic and social benefits for practical applications also becomes a major concern.These two aspects are directly related to the safety and the utility of online social networks,therefore research about them both have theoretical value and practical significance.In the context of rapid development of online social networks,this dissertation conducts in-depth research on the above-mentioned two aspects and proposes some corresponding models and algorithms.The content and main contributions of this dissertation are as follows:*Reputation-Chain Trust Model for Online Social NetworksSince users can propagate trust information in social networks,this dissertation proposes a two-dimensional reputation-chain trust model which integrates trust values and reliable values.According to different application scenarios and users’ personalized requirements,some trust calculation strategies are presented.Due to the openness and anonymity of the Internet,this dissertation introduces several mechanisms of reputation-chain trust model such as local storage mechanism,dynamic update mechanism and trust reporting mechanism.The experiment on the real social network of Epinions shows the flexibility and efficiency of reputation-chain trust model.*Semantic-based Trust ReasoningThis dissertation proposes an innovative semantic-based trust reasoning mechanism. This mechanism exploits the Semantic Web technology to define Epinions domain ontology,then makes use of OWL/RDF language for knowledge representation of Epinions data,and extracts trust-related information from the data.Based on these information,this dissertation defines OWL-based trust reasoning rules.Using these rules,we manage to reason about users’ interested categories,and infer from overgeneralized trust relationships to category-specific trust relationships,and discover implicit trust relationships between users according to their feedback behavior,thus supporting more accurate and efficient trust calculation.*A Reputation-,Content- and Context-based Combined Trust ModelMost existing trust models are based on reputation information,however,only utilizing reputation information for trust computing is far from sufficient.Therefore,this dissertation innovatively puts forward a combined trust model-RCCtrust which integrates reputation,content and context information.RCCtrust utilizes semanticbased trust reasoning mechanism to extract trust-related information from the content and context of the Internet,and integrates users’ similarity in product ratings and feedback behavior to depict the trust degree between pairs of users,thus constructing a edge-weighted combined trust network for trust computing.The experimental results show that RCCtrust model outperforms pure similarity method of traditional collaborative filtering and trust-aware method only utilizing trust relationships both in accuracy and coverage.*Heuristic Information-based Targeted Group DiscoveryThis dissertation proposes a targeted group discovery algorithm based on heuristic information in social networks.According to the features obtained from the analysis of Epinions online social network,combining with users roles and online behavior, we can extract heuristic information for targeted group discovery,thus effectively simplify the algorithm and reduce search space.The experimental results show that the heuristic information-based targeted group discover algorithm manages to efficiently locate influential targeted groups in the online social network and has a widespread application value in social network-based marketing.* Influence Maximization-based Key Persons MiningTraditional methods of mining key persons only considered the structural characteristics of social networks,but neglected the interactions between nodes.Concerning the above disadvantage,this dissertation proposes an influence maximization-based key persons mining algorithm.For the trust network on Epinions,this dissertation utilizes the extracted information from the content and context of Internet,models the influence relationships between users,and identifies key persons in the online social network by solving the influence maximization problem.The experimental results show that our proposed influence maximization-based hill-climbing algorithm outperforms the other methods in different ranges of activation threshold.The advantage of hill-climbing algorithm becomes more prominent especially when the activation threshold is large.更多还原