【作者】 高旻；

【导师】 吴中福；

【作者基本信息】 重庆大学 ， 计算机应用， 2010， 博士

【摘要】 随着Internet的发展,网上的各类资源都以惊人的速度增长着。要及时地在网络的海量信息中发现所需要的资源已经变得越来越困难,用户极需一种推荐系统帮助他们。个性化推荐系统可以帮助用户找到所需信息,能有效留住用户、提高网站的点击率和用户的忠诚度。进而辅助企业达到个性化营销的目的,提升销售量,创造最大的利润。个性化服务概念的兴起,也使得越来越多领域开始重视推荐系统的应用。个性化推荐系统在这种良好的发展趋势的推动和应用前景下,逐渐成为Web智能技术的一个重要研究内容,得到了众多研究者的广泛关注。近年来,协同过滤推荐技术在理论和实践中都得到了快速的发展,但是随着其应用系统规模的进一步扩大,它面临着一系列新的挑战。与此同时,把语用学理论应用于信息系统也逐步成为一个研究热点。语用学是研究符号与解释者之间的关系,研究一定语境条件下的语言和符号产生含义的理解及其产生的效果。本论文将计算语用学的基本思想应用于推荐系统中,包括对推荐系统中个性化情境分析、用户信任等级计算模型、基于情境及信任的推荐方法和算法的应用研究。研究开创性地提出了个性化情境和用户信任等级的概念,从一个新颖的角度解决了协同过滤推荐中的情境缺失和独立性假设问题,提高了推荐系统的推荐质量与抗评分攻击能力;同时,计算语用学在个性化推荐领域的应用研究对语用学本身的发展也具有推动作用。现将论文的主要研究内容和成果概括如下:①对个性化推荐系统目前的总体发展情况进行了综述。探讨了个性化概念的界定,总结并归类了现有的推荐技术,指出其各自的特点、适用范围;在此基础上,对协同过滤算法的目前研究进展进行总结、分类,并指出存在的问题,引出本文的研究意义,为下一步研究奠定理论基础。②对语用学的发展及其与个性化推荐的关系进行了分析。简述了语用学的发展历史,分析了语用学研究与个性化推荐研究的相似性,提出了计算语用学三要素,即溯因推理、信任和情境与个性化推荐的关系,为基于语用学的个性化资源推荐的研究奠定了理论基础。③提出个性化情境的概念、最显著个性化情境因素和多因素个性化情境等概念。把个性化情境引入协同过滤推荐系统,构建了基于个性化情境的推荐方法来解决情境缺失问题。然后提出基于最显著个性化情境因素的协同过滤推荐算法,通过实验证明最显著个性化情境因素对评分预测准确性的提高是有帮助的。进而提出基于多因素个性化情境的推荐方法,采用BP神经网络和RBF神经网络进行情境因素权重的学习,得到基于神经网络的个性化评分预测模型后进行评分预测和推荐,并通过实验验证了算法的有效性,说明多因素个性化情境的推荐更能提高评分预测的准确性。实验还证明径向基神经网络更能提高预测的准确性,基于多因素个性化情境的推荐算法比传统算法有更好的调和平均值和稳定性。④把信任引入协同过滤推荐系统,构建了一种基于信任的协同过滤推荐方法,为解决用户独立性假设问题提供了一种新的思路和方法。在分析信任的定义、性质以及信任与推荐的关系的基础上,提出基于用户兴趣相似性、评分相似性和评分相关性来构建用户关联图的方法,提出基于PageRank用户信任等级的UserRank计算方法,进而提出了基于用户信任等级的协同过滤推荐算法,并通过实验验证了算法的有效性和优越性。实验结果表明:将用户信任等级与经典的推荐算法结合,在不影响预测准确性的前提下可以提高算法防范评分攻击的能力。最后,提出基于情境和信任的综合推荐方法,作为一个综合情境因素和信任因素进行推荐的试探性工作。⑤将以上研究提出的个性化情境分析方法、信任等级计算方法和几种推荐方法用于构建一个基于语用的学习资源个性化推荐系统,列出了系统的体系结构、功能模块设计和结果展示等内容。更多还原

【Abstract】 The volume of information over the Internet is increasing at a tremendous rate. Users are usually confronted with situations in which they have been exposed with too many options to choose from. They need help to explore and to filter out irrelevant information based on their preferences. The corresponding requirement is to describe and capture relevant information for directing users based on specific interests and needs. The supporting systems are identified as personalized recommendation systems. The recommendation systems will also act as a stimulant to further E-Commerce sales growth by converting browsers into buyers, preventing user losing, increasing clicks-ratio, and building customer loyalty. Recently recommendation systems have gradually become an important part in Web Intelligence technologies. Personalized recommendation approaches have gained great momentum both in the commercial and research areas.Recent years, collaborative filtering recommendation technique has gain rapid growth in theory and practice, but along with the scale of the application systems further expanding, it is facing a series of new challenges. At same time, pragmatics theory applied to information systems has also gradually become a research hotspot. This dissertation is focused on the application of the basic ideas of computational pragmatics in recommendation systems. The main research includes: the personalized context analysis, the model for user belief rank computing, recommendation methodologies and algorithms based on context and belief. The research is to solve the context-free, independence-assumption, and shilling-attacks problem in collaborative filtering recommendation approaches from a novel perspective, and to improve the quality of the recommendations. Meanwhile the application of computational pragmatics to personalized recommendation domain will also improve the development of the pragmatics. The main research and contributions in this dissertation are summarized as follows:①The current state-of-the-art methods and techniques of personalized recommendation systems are reviewed. In the dissertation, recommendation techniques are summarized, their characteristics and the scope of application are analyzed. On that basis, the research on collaborative filtering algorithms is summarized and classified, and the problem of the research was analyzed. ②The development of pragmatics and the relationships and similarities between pragmatics and personalized recommendation are analyzed. The relationships between the three key concepts of computational pragmatics and personalized recommendation are further analyzed. The research has set up the theoretical foundation for the research on pragmatics-based personalized recommendation.③Several approaches are proposed to solve the context-free problem by incorporating personalized contextual information into collaborative filtering recommendation systems. Firstly, a collaborative filtering recommendation algorithm based on the most significant personalized contextual parameter is proposed. The experimental results show that our approach is helpful to improve the precision of rating prediction. Secondly, a collaborative filtering recommendation approach based on personalized multi-contextual parameter is proposed. In this approach, a rating prediction model is learned by BP and RBF neural networks. Then the model is used to extrapolate unknown ratings and make recommendations. An experiment is given to evaluate the approach. The experimental results show that the approach is more helpful to improve the precision of rating prediction. The RBF neural network is more helpful than BP is.④A user belief rank based collaborative filtering recommendation approach is proposed. The approach provides a novel idea and method for the independence-assumption problem. Firstly, the definition and characteristics of belief, and the relationships between belief and recommendation is analyzed. Secondly, on that basis, a PageRank-based UserRank (belief rank) algorithm is proposed. Thirdly, and two collaborative filtering recommendation algorithms are proposed based on UserRank and Adjusted Cosine, and UserRank and Slope One. Finally, the experiments show that our approaches provide better recommendation results than Adjusted Cosine and Slope One. At the end, as trial research for the recommendation effectiveness based on the combination of contextual information and belief information, a comprehensive recommendation algorithm is proposed.⑤A pragmatics-based learning resource personalized recommendation system is proposed. In the system, the analysis approach for personalized context, the computation method for user rank, and the personalized context based and user rank based recommendation approaches are applied. The architecture, functional modules, and system interfaces are provided.更多还原