【作者】 肖敏；

【导师】 熊前兴； 钟珞；

【作者基本信息】 武汉理工大学 ， 计算机应用技术， 2009， 博士

【摘要】 随着Internet的不断发展,电子商务系统给商家和客户带来了越来越多的信息,如何及时地在网上的海量信息中发现所需要的信息变得越来越困难。于是电子商务诸多的推荐系统应运而生,推荐技术成为一个研究的热点,引起人们的广泛关注。近年来,电子商务推荐技术在理论和实践中均得到了较快的发展,与此同时,电子商务推荐系统面临着严峻挑战。针对现有电子商务推荐系统存在的问题,本文在电子商务推荐系统中引入领域本体,对电子商务推荐系统中的推荐算法及推荐模型等关键技术进行了深入的研究,以期通过引入领域本体和Web挖掘提高电子商务推荐的准确率和实时性。其主要工作与创新体现在:(1)在探讨和分析比较各种本体构建方法基础上,借鉴软件工程学中的基于软件生命周期模型的方法论,并利用现有的本体构建工具,提出了一种新的基于原型迭代的领域本体构建方法,并构建个性化推荐所需的领域本体。(2)稀疏性问题是协同过滤推荐所面临的最重要问题之一。针对用户评分数据的稀疏性问题,本文提出一个基于领域本体和用户偏好变化的协同过滤推荐算法。利用领域本体中项目的类型及属性计算项目之间的语义相似度,采用KNN(K Nearest Neighbor,K最近邻居)的思想根据用户对项目的评分,预测用户未评分项目的评分,填充用户评分矩阵的缺失值,而后在填充后的用户-项目评分矩阵基础上进行推荐。利用用户的特征因素对用户进行聚类,缩小最近邻居的选择范围。本算法还考虑到用户偏好的变化,引入遗忘函数,根据评价时间调整评分权重。实验结果表明:所提出的算法能够有效地解决稀疏性问题,改善了推荐的质量。(3)传统的Web使用挖掘在个性化推荐过程中没有考虑相关领域的语义知识,不能利用对象的语义进行推荐,从而导致推荐的准确率比较低。针对上述问题,提出一种基于领域本体和Web使用挖掘的个性化推荐模型,将领域本体集成到Web挖掘和个性化推荐中。针对这一模型,本文提出一种基于语义聚类的个性化推荐算法,利用领域本体对Web数据进行预处理,并采用K-Means层次凝聚算法对交易事务进行聚类分析。而后利用各个聚类的质心点矢量来表征每个聚类,生成准确的用户访问偏好和推荐集。(4)提出了多模型的电子商务推荐系统模型,该模型支持非个性化、协同过滤和基于Web使用挖掘的多种推荐。通过挖掘Web使用数据和用户项目-评分数据,分析用户属性信息和用户评分记录等信息,挖掘用户潜在兴趣偏好,在不断的学习中为用户提供准确实时的个性化推荐服务。以此模型为基础,设计并实现了基于电影领域本体的个性化推荐原型系统,验证了其正确性。更多还原

【Abstract】 With the popularization of the Internet, E-Commerce systems bring more and more information for businesses and customers, and it becomes much more difficult for consumers to find services they want in a timely manner from the massive online sources. To address this issue, a variety of recommendation systems were proposed and great attentions have been paid on this new technology, which has become a hotspot in recent researches.Although the recommendation systems in E-Commerce have been very successful in both research and practice, challenging problems still remain. Aimed at solving the main challenges of recommendation systems in E-Commerce, this dissertation attempts to integrate domain ontology with Web usage mining for recommendation personalization and gives a rewarding research on recommendation algorithms and related models in E-Commerce recommendation systems in order to improve the accuracy and the instantaneity of the systems. The main research work and innovative points discussed in this dissertation are as follows:(1) Based on analyzing different means of domain ontology construction, a new method is proposed to construct domain ontology for personalized recommendation.(2) Sparsity is one of the most important issues in collaborative filtering recommendation. To deal with the sparisty of user-item rating, a new collaborative filtering recommendation algorithm which is based on domain ontology and interest drift has been developed.In the new algorithm, the semantic similarity between types and values can be computed according to the domain ontology, the predicting rating of items unfilled by users can be predicted with the semantic similarity and filled according to the K Nearest Neighbors(KNN), then the recommendations can be made using the filled user-item rating matrix. Besides, the user’s features can be used for user clustering to reduce the selection scope of nearest neighbors. Taking the changes in users’ preferences into account, a forgetfulness function f(t) is introduced to adjust the importance of rating considering the rating time. The experiment results show that the new algorithm can solve the sparse problem effectively and has better recommendation quality than the traditional algorithm.(3) Without using the semantics of objects, the traditional Web usage mining in personalized recommendation does not consider relative semantic knowledge, so the accuracy of recommendation is low. In order to solve this problem, a personalized recommendation model which integrating domain ontology with Web usage mining for personalization is presented. An innovative personalized recommendation algorithm based on semantic clustering is devised in this model. An ontology-based vector space model is setting up after the preprocessing on Web usage data with domain ontology. The transaction data are clustered with K-Means Agglomerative Nesting algorithm. The cancroids of clusters can be used to generate user preference and recommendation data sets.(4) A new multiple recommendation prototype system is designed and implemented. The new system can support different multiple recommendation models such as non-personalized recommendation, personalized recommendation for registered users and personalized recommendation for unregistered users. By mining on the Web usage data and user item-rating data, analyzing the user attribute information and user rating records, the model learns the potential interests of the users, and provides instant and accurate personalized recommendation services. The effectiveness of the model is verified through the development of a personalized recommendation prototype system based on a film domain ontology.更多还原