【作者】 孔维梁；

【导师】 杨宗凯； 刘清堂；

【作者基本信息】 华中师范大学 ， 教育技术学， 2013， 博士

【摘要】 随着网络的迅速发展,人类进入了信息社会和网络时代,Internet正为人们提供越来越多的信息和服务。互联网打破了人们传统生活和学习的时空限制,人们可以方便的在互联网上购物,也可以随时随地通过互联网进行学习。但是人们在享受Internet带来的便利的同时,也不得不面对浩如烟海的数据以及大量的垃圾信息,这就带来了著名的“信息过载”和“信息迷航”问题。个性化推荐系统此时应运而生,它能够感知用户的兴趣或需求,而后实现个性化的资源推荐,有效的解决了“信息过载”和“信息迷航”问题。在个性化推荐系统中,协同过滤算法是目前最成功也是应用最广泛的技术之一,但随着用户数的增加以及系统规模的扩大,协同过滤技术面临诸多挑战。本文以协同过滤技术为主要研究目标,从用户使用的角度出发,选择对用户使用有严重影响的关键问题进行深入的探讨和研究,旨在改善协同过滤算法所面临的稀疏性、用户冷启动以及扩展性等问题。针对上述问题,论文在以下几个方面开展理论研究与实践工作：1)针对传统基于项目的协同过滤算法在数据稀疏时的较高预测准确度而低推荐精度的问题,本文提出了基于相关度协同过滤算法(Relation-based Collaborative Filtering, RBCF)。首先,指出传统算法产生候选项目集的不合理,引入关联规则思想,根据项目关联关系生成项目关联矩阵,并使用项目关联矩阵生成候选项目集；然后,针对数据稀疏情况下相似度准确性较差的问题,提出了修正的Pearson相关系数,进一步提高了相似度的准确性；最后,实验证明在评分矩阵相当稀疏的情况下,本文提出的算法能够明显提高推荐质量。同时,该算法大大降低了候选项目集的规模,在一定程度上缓解了算法的扩展性问题。2)详细分析了传统基于项目的协同过滤算法在面对新用户冷启动问题时,待预测项目的近邻数不足等问题,提出了一种基于局部填充的协同过滤算法(Filling-based Collaborative Filtering, FBCF)。算法使用局部最大可能性填充方法扩展新用户的已评分项目集,以便为待预测项目提供较多的近邻。与传统填充算法不同,本文方法不需要额外的用户或项目的属性信息。最后,实验结果证明本文提出的算法能够有效改善新用户冷启动问题。3)随着用户数的增加以及系统规模的扩大,协同过滤算法还面临着严重的扩展性问题。基于专家的协同过滤算法提供了一种新的解决思路,能够在保证相对较高的预测准确度和推荐精度的同时,较为有效地解决扩展性问题。但是,也带来了新的问题——如何有效地选择专家。为此,本文提出了一种基于聚类选择专家的协同过滤算法(Collaborative Filtering Algorithm Incorporated with Cluster-based Expert Selection, CBES)。该算法首先对专家进行了重定义,将专家映射为简单易测量的度量；然后引入聚类方法将项目空间分成不同领域,从每个领域对应用户中选择出代表组成专家集；最后,实验结果表明在预测准确度和推荐精度上,本文提出的算法取得了较好的效果。4)在前文所作研究的基础上,本文探讨了面向适应学习服务的协同过滤系统研究与实现,文章设计并实现了基于村镇教育平台的协同过滤推荐系统(Recommendation System for Country Education Platform, CEPRS)。系统综合实现了本文提出的算法,为在线学习用户提供了个性化服务,使得在线教育平台在用户中心化上前进了一大步。文章介绍了系统的体系结构和主要功能模块,并对算法的实现进行了简要说明,另外,该系统具有良好的可移植性、可维护性及开放式架构的特征。更多还原

【Abstract】 With the rapid development of Internet, the human race has entered the information society and the network era. Internet provides people with more and more information and services and it has broken the limit of space and time of traditional life and learning. People can shop on the Internet conveniently and study via Internet whenever and wherever. However, people have to face enormous data and useless information when they enjoy the convenience brought by Internet. This is the famous "information overload" problem and "information confusion" problem. Personalized recommender system has emerged in response to this challenge and it provides users with personalized recommendations of items which are likely to fit their needs.Collaborative filtering algorithm is a core technology of personalized recommender systems, and it is also one of the most widely used and successful technologies. But with the expansion of system, collaborative filtering algorithm is facing many challenges. This paper concentrates on some key problems faced by collaborative filtering algorithm, which are the sparse problem, the new user cold-start problem and the scalability problem. This paper does some research on the following aspects:1) In order to solve the problem of low recommendation precision faced by item-based collaborative filtering algorithm when the user-item rating matrix is sparse, this paper proposes a relation-based collaborative filtering (RBCF). This algorithm points out the shortage of traditional item-based collaborative filtering algorithm when generating candidate item set. Then it proposes to generate candidate item set using item correlation matrix which is calculated according to the thought of association rules. At the same time, in order to alleviate the inaccurate of similarities when the user-item rating matrix is sparse, this algorithm adjusts the Pearson correlation. Finally, experiments show that the proposed algorithm could greatly improve the performance of recommendation. In addition, the propose method can also reduce the size of candidate item set obviously, which alleviates the scalability problem of collaborative filtering algorithm.2) This paper points out the shortages of the traditional collaborative filtering algorithm when facing the new user cold-start problem, and then proposes a filling-based collaborative filtering altorithm (FBCF) basing on the traditional item-based collaborative filtering altorithm. The proposed algorithm proposes a partial filling method to extend the new user’s rated item set with a high efficiency. Besides alleviating the new user cold-start problem, another advantage of the proposed algorithm lies in with no need for additional information except the basic user-item rating matrix. In order to evaluate the proposed algorithm, experiments based on the well-known dataset were conducted to compare the FBCF algorithm and other benchmark algorithms, and the results show that the FBCF algorithm has an obviously better performance.3) With the expansion of the system and the increase of the number of users, the collaborative filtering algorithm not only faces the sparse problem but also the scalability problem. The expert-based collaborative filtering algorithm shows a new solution to this problem, and it can solve the scalability problem effectively while maintaining a relatively high prediction accuracy and precision. But it brings another problem—how to select the experts effectively. Therefore, this paper puts forward a new collaborative filtering algorithm incorporated with cluster-based expert selection (CBES). In the algorithm, this paper firstly redefines the expert and maps it to a simple and easy measurement metrics. Then this paper clusters the users with the same interests as a group, and selects representative experts from each group. Finally, this paper compared the proposed algorithm to the traditional user-based collaborative filtering and the algorithm proposed by Xavier, and the results show that the proposed algorithm can get well performance on prediction accuracy and recommendation precision.4) Based on the above research, this paper discusses research and implementation of recommender system for adaptive learning services. This paper designs and implements a recommender system for country education platform (CEPRS) based on rural education platform. The system implements the proposed algorithms, provides personalized service for online learning and makes online education platform get a big step forward on user centric. This paper introduces the architecture and the main function modules of the system, and gives out a brief explanation to the implementations of the proposed algorithms. This system has good portability and maintainability characteristics.更多还原