【作者】 李晓建；

【导师】 陈世鸿；

【作者基本信息】 武汉大学 ， 软件工程， 2010， 博士

【摘要】 随着信息时代的来临以及互联网技术的发展和普及,人们对自身周边信息的重视度以及信息获取的要求越来越高。互联网已经成为信息领域的百科全书。但是,互联网作为一个信息空间具有开放性、异构性、分布式、管理非一致性等特性,而且具有极快的更新和进化速度,对于用户而言,如此巨大、无序的互联网信息空间中,使其真正感兴趣的信息实为沧海一粒。面对如此海量的信息,如何快速、准确地找到符合用户要求的信息,引起了广大学者的关注。通过大量的研究,有关学者提出可以通过构建个性化资源推荐系统,基于用户的资源偏好帮助用户在系统中找到最符合用户要求的资源。目前用于构建上述资源推荐系统的主要技术有很多,但在实际的应用环境中,资源推荐的准确性一直无法得到有效地提高,其主要原因包括：1)缺乏有效地资源描述模型；2)基于当前算法和技术生成的用户偏好模型,往往因为系统采集的用户偏好数据不足而无法正确的反应用户偏好；3)资源推荐的算法规模因系统中资源数据量的庞大而得不到有效地控制,过大的系统开销制约了系统的高效运行。本文将基于本体的语义技术引入资源推荐系统中,尝试从各方面解决上述问题,主要内容包括：(1)基于本体的资源描述建模使用多继承式级层结构的本体构建资源描述模型,以更加有效地提取资源间的关系；从本体结构中抽取描述概念的语义关联和背景知识,降低用户对资源的认知要求；同时利用语义信息,系统可以使用本文提出的算法从已有的信息中推理出新的信息,用来补足用户偏好数据不足的缺陷,使得偏好获取的过程更加简单和高效。(2)用户偏好的推理机制基于资源描述本体模型中描述的用户概念偏好和抽取自本体结构的语义信息,通过概念之间语义关联的计算,找到语义相似度最近的概念,实现用户偏好在概念之间的语义推理传递,解决用户偏好数据采集不足时,用户模型的正确有效构建之问题,以提高资源推荐的准确性。(3)概念间语义相似度度量算法基于对用户评分推理过程的研究,由本体结构中抽取出的概念间语义信息,通过对比用户对概念的评分描述概念间共享的特征,使用对概念推理传递评分的量化计算,描述概念间共享特征的程度,以衡量概念间的相似关系,当概念间共享的特征越多则概念越相似。(4)本体的自动构建及个性化计算在实际的应用环境中,通常没有存在的领域本体,本文通过研究资源推荐系统中用户与资源的关系,提出使用聚簇算法自动创建本体的方法,基于聚簇算法中不同条件函数的应用,构建若干个独立的简单树型结构本体,并依据缩小用户偏好推理过程中可能的推理偏差原则,寻找最符合用户偏好的资源描述本体,有效的提高个性化资源推荐的准确度；同时通过对经典聚簇算法的扩展给出构建具有多继承级层结构的本体算法,进一步改进推荐系统的准确程度。(5)资源推荐语义模型综合所有创新的算法和技术,给出了一个基于用户偏好的语义个性化资源推荐模型框架。并通过具体实验证明了使用该模型能够比经典的协同过滤模型取得更好的资源推荐效果。更多还原

【Abstract】 With the advent of the information age and the development of Internet technology, it has become increasingly demanding that people focus on accessing peripheral information. However, the internet as an information encyclopedia not only has characteristics including open, heterogeneous, distributed and non-uniformity, but also updates and evolves extremely fast. In the disordered Internet, Information that users really interested in is rare. Faced with this flood of information, how to quickly and accurately meet user requirements to find the information has aroused the attention of scholars. Through investigation and study, researchers have proposed measures by building personalized recommendation system, to help users find the best resources based on the user’s preferences and requirements.Currently, there are a lot of technologies for building the recommendation system. However in actual application environments, the accuracy of the system has not been able to effectively improve. Main reasons are:1) Lacking of effective resource description model; 2) User preference model which is built based on current algorithms can not fully express the user’s preference while the user preference data collected by the system are usually not correct and sufficient; 3) Due to the huge amount of resources, the complexity of the algorithm can not be controlled well which will restrict the efficient of the system. This dissertation try to utilize semantic technology to solve above problems, the main contribution is described as follows:(1) Resource description model based on OntologyUse ontology with multi-hierarchical structure to build resource description model and extract relationship between resources; Extract semantic relation and background knowledge from ontology structure to reduce the user’s cognitive burden; By using those semantic information and algorithms proposed by this dissertation, the system can infer new information to complete the user preference model and make the preference elicitation process more simple and efficient. (2) Inference mechanism of the user preferenceBased on user preference description and abstracted semantic information from the resource description model, compute the semantic similarity between concepts and find the most close concept to enable the semantic propagation of the user preference between concepts in order to solve the problem of being not able to fully express the user’s preference as the user preference data collected by the system are usually not correct and sufficient.(3) Semantic similarity metric algorithmAccording to the preference inference process, based on extracted semantic information from ontology and common features of the concepts, compute the amount of propagated user preference to describe the degree of common features between concepts and express the similarity between concepts.(4) Ontology learning algorithm and personalization computationIn practical application environments, there exists no domain ontology. This dissertation proposes the tow ontology learning algorithms to automatically construct the ontology by utilize clustering algorithm and investigate the relation between user preference and resource. The first one allows system to build a set of distinct hierarchical ontologies, while the second extends classical agglomerative clustering and builds a multi-hierarchical ontology. An algorithm for selecting which ontology to use based on the user’s preferences is also presented to increase the accuracy of resource recommending.(5) Ontology based semantic resource recommendation modelBy integrating all algorithms and techniques brought, this dissertation proposes an ontology based semantic resource recommendation model to help users find the best resources based on the user’s preferences and requirements. Experiments are carried out to prove that our model performs better than traditional models in resource recommendation scenario.更多还原