【作者】 郭艳红；

【导师】 邓贵仕；

【作者基本信息】 大连理工大学 ， 管理科学与工程， 2008， 博士

【摘要】 随着网络技术的应用和普及、电子商务的迅猛发展,越来越多的信息充斥在网络之上。如何在众多的资源中找到适合自己需求的信息,成为众多学者、专家和网络用户关心的核心问题之一。推荐系统在这样的背景下应运而生。协同过滤技术是推荐系统(Recommender System)最为核心的技术之一,也是目前应用最为广泛和成功的技术。本文以推荐系统的协同过滤算法为研究目标,旨在解决协同过滤算法在应用中所遇到的稀疏性问题、冷启动问题、用户的信任问题等关键问题。针对推荐系统的协同过滤算法,本文在以下几个方面作了相应的理论研究和应用工作:1、对目前推荐系统的总体发展进行了综述,从心理学和认知科学的角度探讨了个性概念的界定,试图从信息科学以外的角度探讨推荐系统构建是否可行及其意义;总结归类了现有的推荐技术,指出其各自的特点、适用范围;在此基础上对推荐系统实现的体系结构和模块进行了概括,为进一步的应用工作提供理论指导;最后对协同过滤算法目前的研究进展进行总结、分类,指出存在的问题,为下一步的研究奠定理论基础。2、提出了稀疏矩阵下的一种基于最近邻居评价的改进的协同过滤算法。首先通过余弦相似度计算公式,为目标用户选取最近邻居集,然后根据每个最近邻居对项目的评价与用户间的相似度,产生一个对项目的预测值,所有最近邻用户对项目的加权预测值就构成了一个虚拟的最近邻评价矩阵。由于这个矩阵体现了与目标评价在空间上的相似性,从而将目标评价的预测问题转化到一个最近邻评价矩阵上进行预测计算。和历史评价矩阵相比,虚拟最近邻评价矩阵不但规模比较小,而且包含了用户维上和项目维上对目标评价最有价值的信息。最后再根据虚拟的最近邻居评价矩阵,进行加权均值预测。实验表明,本文提出的针对稀疏矩阵的改进算法,在精度上优于传统算法,尤其在最近邻居个数较少的情况下,精度有较大提高。3、提出基于项目关键词预测与协同过滤相结合的混合推荐算法。分析了在系统中项目的内容信息不够丰富的情况下,如何应用基于项目关键词预测与协同过滤技术相结合的问题。首先把项目的关键词进行二进制代码表示,以达到对项目内容进行形式化描述的目的,然后通过winnow算法,对用户的评价进行初步预测,得到用户的预测矩阵;为了确保用户评价矩阵在空间上同目标评价的相似性,构造了两个约束参数——用户评价密度α_i和预测精度β_i,通过参数的设定和遴选,在满足约束条件的评价数据集上应用协同过滤算法。实验表明,加入约束条件的混合推荐算法远优于传统的协同过滤算法以及没有任何约束条件的混合推荐算法。4、提出了把信任引入协同过滤推荐系统的构想,构建了一种基于信任的协同过滤推荐算法,在对信任进行了形式化的定义和描述的基础上,构造了协同过滤推荐系统中的两种信任模型——局部信任(local trust)和全局信任(global trust),并分别指出两种信任的区别,确定了各自的影响因素,通过这两种可计算的信任模型可以对系统中的用户的信任程度进行不同范围内的度量;进而,提出了一种基于信任因子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。最后的实验同样分析了两种信任的分布特性,通过与相似度的分布的对比,可以得出结论:在推荐系统中对用户信任的研究是有意义的,信任是同相似度不同的对最后的推荐产生影响的重要因素之一。5、构建了基于事例推理(CBR)的推荐系统框架模型。对基于事例的推理和协同过滤的推荐过程进行了比较,指出异同,进而把基于事例的推理的过程与推荐系统相结合,应用基于事例的推理过程更好地更新用户的档案,跟踪用户的兴趣变化,提高系统的学习能力。在框架模型中,本文对推荐系统的各部分作了充分的总结和说明,为今后的进一步理论和实践研究奠定基础,同时实现了一个以电影推荐为例的电影推荐系统。通过上述的研究工作,从一定程度上解决了推荐系统的协同过滤算法所遇到的稀疏性问题、冷启动问题、信任问题,从而从一定程度上推动协同过滤算法的理论研究和应用研究的进展。更多还原

【Abstract】 With the fast development of Internet and applications of E-Commerce, more and more information swirles in the net. To get the right information from the information sea has become one of the key issues nowadays for the researchers, experts and the Internet users. Personalized Recommender Systems emerge under the background of this, which becomes the research focus in the domestic and overseas.Collaborative filtering algorithm is the most key technology in the personalized recommender systems, which has got the most success and wild applications. This dissertation takes collaborative filtering algorithm of personalized recommender system as the research project to deal with the sparsity problem, cold start problem and trust problem, etc. Research work are taken as following:(1) Review the research development of personalized recommender system and discuss the concept of personal from the angle of recognization and phsycology to give some advice for the configuration of personalized recommender systems; give a general analysis of recommender technologies and indicate their individual characters and application fields; analyze the system constructure and modules to give some instructions for applications; at last, give a division of the collaborative filtering and indicate their challenges for research work.(2) Put forward an improved collaborative filtering algorithms based on nearest neighbor rating matrix for the sparsity problem of collaborative filtering algorithm. First, get the nearest neighbor rating matrix through Cosine similarity calculation metrix and produce a prediction for items according to similarity between users and neighbors’ ratings. Then all nearest neighbors’ prediction ratings forms a virtual nearest neighbor rating matrix. This matrix takes on a similarity with the active user rating and the problem of prediction for the active user could be transfer to the matrix. Compared with history rating matrix, virtual nearest neighbors’ rating matrix has small scale and contains the most useful information. At last, a prediction based on weights of similarity between users and ratings is produced. Experiments improved the improved algorithm this paper advanced and especially when the rating matrix is very sparse, prediction accuracy is much better.(3) Advanced a hybrid recommender algorithm based on the combination of collaborative filtering and item keywords based prediction. Analyze the combination problem of item keywords based prediction and collaborative filtering when item keywords information is not adequate. Through the content abstraction of the items, items in the recommender system are represented by 0 and 1. After this users’ prediction for the items are got by using winnow algorithm. To guarantee the accuracy of the prediction matrix, two constraint parametersα_i (which indicate the rating number of each user) andβ_i (number of the prediction which is accurate ) are used to filter the prediction. Only those users’ predictionwhich is accurate could be into next step filtering——collaborative filtering. Experimentimproved that the hybrid recommender algorithm with the constraint parameter is much superior to the traditional collaborative filtering algorithm and the algorithm without the constraint parameters. The hybrid algorithm this dissertation put forward improved the cold start problem from certain extent.(4) Put forward an assumption that introduce social trust into recommender systems andconstruct a collaborative filtering algorithm based on trust. Configuate two trust models——global trust and local trust under the foundation of formatlate description of the trust and then indicate their effective parameters. Through this two calculatable trust model, user trust could be measured and a collaborative filtering algorithm based on trust is then advanced. The expeiriment proved the superiority of the algorithm. In the experiment the distribution of the trust of two trust model and the distribution of similarity are also be demonstrated which could be safely concluded that trust is a parameter that affect the final recommendation which is very different between the similarity. So the assumption advanced at first is very meaningful.(5) Construct a general model of the personalized recommender system based on CBR. Comparing the difference between CBR and CF, the similarity and difference is generalized. Then CBR and CF are combined together which using CBR to improve the learning ability of the personalized recommender system. A personalized film recommender prototype system is also developed at last which could prove the models and algorithms this dissertation put forward.更多还原