【作者】 赵伟；

【导师】 郭拯危；

【作者基本信息】 河南大学 ， 应用数学， 2007， 硕士

【摘要】 随着互联网的普及和网络技术的不断发展,电子商务因其成本低廉、快捷、不受时空限制等优点在全球范围内得到普及和发展。而在这种虚拟环境下,商家所提供的产品种类和数量非常多,而且,从现实经验来看,用户的需求经常是不明确的、模糊的,可能会对某类产品有着潜在的需求,但并不清楚什么产品能满足自己的模糊需求,所以,如何向用户进行个性化产品推荐,将电子商务网站的浏览者转变为购买者,提高电子商务网站的交叉销售能力以及客户对电子商务网站的忠诚度,使得企业获得尽可能多的效益,成为电子商务的一个重要研究内容。推荐系统就是在这样的背景下与电子商务结合在一起的。协同过滤是目前在电子商务推荐系统中应用最早和最为成功的个性化推荐技术。但是,随着站点结构、内容复杂度、产品数量、产品种类和用户人数的不断增加,推荐系统中的协同过滤技术发展面临着诸如推荐质量不高、扩展性差等严峻的挑战。面对这些挑战,针对如何提高对用户推荐信息的质量、提高协同过滤算法的可扩展性等问题,国内外进行了很多研究,其成果也在实际中得到了一定的应用。论文首先简要介绍了电子商务推荐系统产生的背景。论文对推荐系统的概念、作用,以及常用的方法等进行了详细的阐述。在推荐技术中,协同过滤技术不仅为推荐系统提高服务质量提供了新的思路,而且该技术在许多商业网站上也得到了广泛地、成功地应用。对协同过滤基本思想、出发点,实现,以及协同过滤的两个方向作了全面介绍。论文还着重分析了协同过滤在推荐系统中应用时所面临的问题,以及现有的解决方法。针对这些问题提出了新的解决方法。该方法是在原有的基于用户的协同过滤和基于项的协同过滤两种算法基础上,引入了概率融合框架融合了前面的两种算法,这样,一方面增加了进行推荐时可用的数据,减少了由于数据稀疏性的影响,在一定程度上提高了推荐质量。另一方面对于协同过滤算法中出现的数据的极端稀疏性问题,引入了BP神经网络的方法,对未评分数据进行了预测,降低了数据的稀疏度,在一定程度上也提高了算法的精确度。另外,针对以往协同过滤算法中未考虑和利用项分类信息,使得推荐时缺少个性,难以适应目前电子商务系统日趋多样性和个性化的趋势的问题,引入了面向场景的方法,这样,既解决了上述问题,又减少了BP神经网络预测时的计算量和相似用户计算的复杂度,对推荐质量的提高也有一定的效果。论文还对提出的算法在标准数据集上给出了实验结果,并将其与其他算法的相关性能进行了比较分析。最后,对本文研究进行了全面的总结,指出了研究中存在的不足,展望了未来进一步研究的方向。更多还原

【Abstract】 With the widespread of network and the development of network technologies, e-commerce has been popularized and development in a global context because of its advantages such as low cost, fast, free from constraints of time and space. But in this virtual enviorment, more and more types and amounts of products are provided by businesses, and from a practical experience, the requirements of users are often unclear, vague. They may have potential demand for certain products, but it is not clear what products would meet their fuzzy needs, so how to give the customers personalized recommendation about the products, how to turn the e-commerce website browsers to buyers ,how to enhance the ability of cross-sell of e-commerce website and the customers loyalty on e-commerce websites, making enterprises maximize their profits have become an important isuue on e-commerce. Recommerder system is combined with e-commerce in such enviroment. Collaborative filtering is the earlist and most successive personalized recommendation technology in the e-commerce recommender systems. Yet, with the continues increment of the structure of website,the complexity of contents, the amounts of products,the types of products and the amount of customers, the development of collaborative filtering technology of recommerder systems faces serious challenges such as poor recommendation quality and scalability. Facing these challenges, on the issues of how to improve the quality of users recommendation information and how to enhance the scalability of collaborative filtering algorithms, some researches have been done in domestic and foreign, and the research results have been successfully applied in practice.This paper briefly introduced background of the e-commerce recommender system at first. It illustrates the recommender system concept, effect, and popular methods in detail. In the recommendation technology, collaborative filetering not only provides new ideas for improving serve quality of recommender systems, but also is applied widly and successfully in many commencial website. It introduces colloaborative filtering basic ideas, theory start, implementation and its two directions. Also it gives emphasis to analyzing the probelmes which collaborative filtering is facing when its applied in recommender systems and existing improved methods. The paper proposes new methods to solve those problems. Based on the user-based and item-based collaborative filering algorithms, the new method introduces probabilistic fusion framework to fuse these two algorithms. Thus, on one hand, it increases the available data during recommending, reduces the effectness of data sparsity, and enhance the recommendation quality to a certain extent. On the other hand, for data extrem sparsity problem, BP neural network is introduced to predict the values of the null ratings, alleviate this issue, and also improves the recommender systems precision to some extent. Furthermore, because of not concerning and use of items classification information in old algorithms, it makes the recommendation lack of personality so that it is difficult to adapt to the trend of e-commerce growing diversity and individuality of the current systems. For this issue, a scene-oriented approach is introduced. Thus, it not only solves the former problems, but also reduces the amount of computation when prediction using BP neural network and similar user computing complexity. It has some positive impact on recommendation quality. It also gives experiment results for proposed algorithnms in standard data sets, and the performance between the new method and the old one is compared and analyzed. Finally, we summarize on the paper, point out defects and the directions that will be further studied in the future.更多还原