【作者】 高翔；

【导师】 马静；

【作者基本信息】 南京航空航天大学 ， 管理科学与工程， 2011， 硕士

【摘要】 随着互联网的普及,人们在海量的各类信息资源中很难找到自己需要的信息,往往会陷入信息过载的困境。在这样一个大背景下,个性化推荐系统应运而生。个性化推荐系统可以根据已有的用户信息,向目标用户推荐其可能感兴趣的信息,可以帮助用户更方便、更快速地找到所需的信息。而协同过滤技术是个性化推荐系统中使用最广泛和应用最成功的技术之一,对它进行研究在理论和应用上都有非常重要的意义。本文首先对个性化推荐系统的基本知识以及主流推荐技术进行研究,然后分析了协同过滤推荐技术中应用最广泛的基于用户和基于项目的协同过滤算法,接着引出这两种协同过滤推荐算法的瓶颈问题:数据稀疏性、冷启动、扩展性问题,并总结了几种解决方法。本文分别从两个方面针对基于用户协同过滤算法进行改进,针对数据稀疏性的瓶颈问题提出了改进算法ICF-AD:其一是通过关联分析对用户-项目评分矩阵进行项目之间关联关系的挖掘,并利用关联关联对用户未评分项目进行预测评分,以实现对用户-项目评分矩阵补值的目的,降低数据稀疏性;其二是提出“领域”的概念,在领域内计算用户之间的相似性,既保证了相似度计算的可靠性,又降低了算法的复杂度,提高了算法性能。本文不仅对改进算法ICF-AD进行了理论分析,阐述其可行性,而且给出了改进算法在标准数据集上的实验结果,并对ICF-AD与其他两种基于用户的协同过滤算法进行了相关性能的比较分析。实验表明,改进的算法ICF-AD不仅能够有效缓解数据稀疏性问题,并且提高了算法性能,推荐质量得到了增强。更多还原

【Abstract】 With the popularity of the Internet and e-commerce application, it is difficult for them to find their needed products within a mass of product information.Therefore, the recommendation system in e-commerce came into being. Recommendation systems were proposed to provide the target user with information they interested in based on the existing information, so users can be more convenient to find the information they need. Collaborative filtering is one of the most widely used and successful methods for recommendation. It’s significant to study it.We introduced the basic knowledge of recommendation system and main recommendation skill detailedly in this thesis, and analyze the most popular user-based collaborative filtering and the item-based collaborative filtering in the collaborative filtering algorithms, then lead the bottlenecks problem of collaborative filtering: data sparsity, cold start and scalability, summarizing some solutions. We proposed two improved methods for sparse problem from two different parts, and bring the improved collaborative filtering. One of the methods is to mine the association of these items through the rating matrix, then complement the matrix with these associations; another one is bringing the concept of“domain”and calculating the similarity of two users in the domain. It not only ensure that the reliability of the space but also reduce the algorithm complexity and time complexity. We theoretically detailed analyze the new methods and prove their feasibility. Then the experimental results that the new method is implemented with the benchmark experimental data set are given, the performance between the new methods and the old methods is compared and analyzed. The experiments show that the improved methods can alleviate the problem of sparsity effectively and scalability, improve the quality of recommendation and efficiency.更多还原