【作者】 陶红亮；

【导师】 王明文；

【作者基本信息】 江西师范大学 ， 计算机应用技术， 2007， 硕士

【摘要】 随着Internet的普及和网上贸易的迅速发展,电子商务系统已成为人们网上购物的主要平台。在给用户提供更多选择和方便的同时,其结构也变得更加复杂和庞大,用户经常会迷失在大量的商品信息空间中,无法迅速找到自己所需的商品。这种情况下,推荐系统应运而生,它在电子商务系统中具有良好的发展空间和应用前景,已成为电子商务领域研究的一个重要内容。推荐系统是根据用户已有的评价信息或历史记录实现个性化服务的系统,向用户推荐其感兴趣的商品或信息。人们将数据挖掘中的各种技术应用于推荐系统的研究,取得了很多成果并且开发出一些优秀的推荐系统,极大地促进了推荐技术的发展。但是,现代电子商务系统中用户和项目数以万计,而且发展非常迅速,使得推荐系统面临一些困难和挑战,主要包括:推荐准确度、实时性要求、数据稀疏问题和可扩展性问题。协同过滤是目前应用最广泛的推荐技术。针对推荐系统面临的数据稀疏问题,本文提出了两种改进的基于聚类的协同过滤算法:基于项目平滑和聚类的方法,双向聚类迭代的方法。将用户和项目分别聚类,在与目标项目最相似的前若干个聚类簇中搜索它的最近邻居,可以缩小搜索邻居的范围,提高推荐算法的实时响应速度。基于项目平滑和聚类的方法中,利用聚类信息,对用户未评分的项目做平滑处理,使得用户-项目矩阵变得稠密,然后在项目聚类中寻找目标项目的最近邻居,在一定程度上可以解决数据稀疏对推荐精度的影响。双向聚类迭代的方法中,引入二部图概念来表示用户和项目之间的关联关系,对初始得到的用户聚类和项目聚类进行交叉迭代调整,使聚类簇达到较稳定的状态。调整后聚类簇的内聚性更强,类之间的区分度更大,这样也可以在一定程度上解决数据稀疏问题的影响,提高推荐的准确度。本文的实验采用标准的MovieLen数据集,用K-mean聚类法对用户和项目进行聚类,观察在不同聚类数目的情况下,类内距与类间距之比。然后考察本文提出的两种方法的MAE值(平均绝对偏差),并与其他一些传统的协同过滤算法的MAE结果相比较,来验证算法的有效性。实验结果表明,相比于其他一些传统的方法,两种新方法在效果和效率上都表现得较好,能够有效解决数据稀疏问题,提高推荐系统的实时响应速度。更多还原

【Abstract】 With the wide spread use of the Internet and the development of commerce on it, E-Commerce System becomes the most important platform to users who have commerce online. It provides more choice and convenience to users than before, but also brings some difficulties to users and the structure of E-Commerce System becomes more and more complex. Customers always confront with too many items and easy lost in it, they are difficult to find the products and services they wanted rapidly. Recommendation Systems emerge as the times require, it becomes the hotspot research field in E-Commerce, it will have good development and application perspective in the future.Recommendation System works based on the users’rating information or history records of users, it recommends the items or information to users which they are possible interested in. Many researchers apply some technologies of Data Mining to the research of Recommendation System, achieved great successful and developed some excellent Recommendation Systems, it greatly promote the development of recommendation technology. However, there are thousands of users and items in modern E-Commerce System, and with high increasing speed. So Recommendation System is confronting with some difficulties and challenges, such as: recommend accuracy, real-time requirement, data sparsity and scalable problem.In this dissertation, we propose two improved clustering-based collaborative filtering algorithms to solve the data sparsity problem. There are: Item Smoothing and Clustering Algorithm, Bidirectional Clustering Iterative Algorithm. By clustering users and items respectively, the nearest neighbors of target item can be found in the several most similar item clusters. In this way, we can reduce the search space, and improve the real-time performance of recommendation algorithm.In the Item Smoothing and Clustering Algorithm, based on the clusters’information we apply the smoothing strategy to the unseen rating data. The user-item matrix becomes densely, and then we find the nearest neighbors of target item in the item clusters. In this way we can solve the data sparsity problem in a certain extent.In the Bidirectional Clustering Iterative Algorithm, we use a bipartite graph to describe the associate relation between users and items. Using the crossing iterative algorithm to adjust initial user and item clusters to reach a steadily status. In this way, can also solve the data sparsity problem and improve the recommend accuracy, as the clusters become more accurate and more steadily.We conduct a series of experiments to examine the effectiveness of our new algorithms; these experiments are all based on the MovieLens dataset. First use K-mean algorithm to cluster users and items, observe the performance under different cluster number. Comparing the MAE (Mean Absolute Error) of the two new algorithms to several traditional collaborative filtering algorithms, such as: User-based, Item-based and Item-cluster collaborative filtering algorithm. The experimental results show that new algorithms could effectively alleviate the data sparsity problem and reduce the MAE, and improve the real-time performance of recommendation system.更多还原