融合用户属性和兴趣对比度的协同过滤个性化推荐研究

发布时间：2018-02-20 04:40

本文关键词： 个性化推荐协同过滤用户属性兴趣对比度　出处：《华中师范大学》2014年硕士论文　论文类型：学位论文

【摘要】：为解决信息过载问题和应对用户对个性化服务的需求,个性化推荐技术应运而生,本文希望通过对个性化推荐的优化与创新,让用户能够更快更精准的找到自己想要的资源。在众多个性化推荐技术中,协同过滤算法是当下研究的热门。因为其算法的应用范围最广泛,发展时间最长,算法最成熟。协同过滤推荐主要是根据存贮在系统数据库中用户历史消费及评分数据,来分析用户的兴趣,预测用户未来可能消费什么样的产品,从而对其实施个性化推荐。以往的协同过滤算法研究,主要是以用户评分矩阵为基础,进行用户偏好的感知,以用户打分的相似性来判断用户之间兴趣的相似性。随着算法的发展,特别是可扩展性问题、冷启动问题等算法瓶颈的出现,纯粹依赖评分矩阵数据来寻找最近邻,就显得力不从心。因此,必须寻找其他有效的用户偏好数据来源。本文对用户偏好的感知方法进行改进,引入了用户属性信息这一重要的偏好感知数据源,与评分矩阵共同构成用户偏好感知的数据基础。用户属性作为描述用户个体特征的重要信息,不同的属性可以将用户划分到不同类别的群体当中,这些用户群可能存在一定的兴趣偏好相似性,将这些特定用户群的共同兴趣找出来,作为产生推荐的基础。本文定义了一个新的衡量用户兴趣偏好的参数,即：兴趣对比度。在此基础上,提出了一个融合用户属性和兴趣对比度的协同过滤个性化推荐算法,该算法以用户属性组合为约束,结合兴趣对比度共同产生待推荐集合,经过整理删选后形成最后的推荐列表。本文提出的新算法,将克服传统协同过滤的可扩展性问题作为改进的目标。在新算法的整个流程设计中,不依赖传统的用户相似性计算来寻找最近邻。因此,当用户和项目快速增长时,不会出现算法复杂度急剧上升的情况,实验证明,融合用户属性和兴趣对比度的协同过滤推荐算法,能够在保证推荐实时性的前提下,达到满意的推荐质量,是一种灵活高效的推荐方案,更重要是提供了一种新的推荐思路。此外,本文还对不同属性组合下的推荐效率进行了系统分析,为该领域的相关研究奠定了一定的基础。
[Abstract]:In order to solve the problem of information overload and to meet the needs of users for personalized service, personalized recommendation technology emerges as the times require. This paper hopes to optimize and innovate personalized recommendation. Among the many personalized recommendation technologies, collaborative filtering algorithm is a hot research topic, because it has the most extensive application and the longest development time. The most mature algorithm is to analyze the interests of users and predict what kind of products they may consume in the future according to the historical consumption and scoring data stored in the system database. In order to implement personalized recommendation, the previous collaborative filtering algorithms are mainly based on the user score matrix, the perception of user preferences, With the development of the algorithm, especially the problem of scalability, cold start problem and other bottlenecks, we rely solely on the score matrix data to find the nearest neighbor. Therefore, we must find other effective sources of user preference data. In this paper, we improve the perception method of user preference and introduce user attribute information as an important data source of preference perception. Together with the score matrix, it forms the data base of user preference perception. As an important information describing the individual characteristics of users, different attributes can divide users into different groups. These user groups may have a certain similarity of interest preferences. The common interests of these specific user groups can be found as the basis for producing recommendations. In this paper, a new parameter to measure user interest preference is defined. That is: interest contrast. On this basis, a collaborative filtering personalized recommendation algorithm combining user attributes and interest contrast is proposed. After sorting and deleting, the final recommendation list is formed. The new algorithm proposed in this paper aims to overcome the scalability problem of traditional collaborative filtering. In the whole process design of the new algorithm, We do not rely on the traditional user similarity calculation to find the nearest neighbor. Therefore, when the user and project grow rapidly, the algorithm complexity will not rise sharply. The collaborative filtering recommendation algorithm which combines user attributes and interest contrast can achieve satisfactory recommendation quality on the premise of ensuring real-time recommendation. It is a flexible and efficient recommendation scheme. More importantly, it provides a new way of recommendation. In addition, this paper also makes a systematic analysis of the efficiency of recommendation under different attribute combinations, which lays a foundation for the related research in this field.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：F224;F713.36

【参考文献】