【作者】 吴婷；

【导师】 熊前兴；

【作者基本信息】 武汉理工大学 ， 计算机应用技术， 2009， 硕士

【摘要】 近年来,电子商务个性化推荐系统在网络上获得了普遍的成功,协同过滤是其中应用最为广泛的个性化推荐技术。但是当前的电子商务推荐系统在实际运用中还相当不成熟,仍然存在许多问题,如推荐质量受到稀疏的用户评价数据的严重影响,系统的可扩展性能差,推荐缺乏多样性无法涵盖用户的完整兴趣。同时,网上有效信息的数量和商品的种类的急速增长对推荐系统提出了严峻挑战。本文研究了个性化推荐系统及其主要的推荐技术,特别是协同过滤技术,包括基于用户的协同过滤技术和基于项目的协同过滤技术。本文所做的主要工作及创新体现在下面的四个方面:本文在协同过滤算法的计算用户间相似度阶段,提出了一种基于用户兴趣变化的协同过滤的改进算法。算法考虑了用户评价时间的影响,改进了传统的用户间相似度的计算方法,从而得到最有效的目标用户最近邻居。另外,在推荐系统中寻找目标用户最有效邻居方面,本文利用用户的属性特征对用户进行聚类。先找到目标用户所在的聚类簇,然后在这个聚类簇中利用改进的用户相似度量方法寻找目标用户的最近邻居。在预测阶段,本文利用能使改进算法达到最低MAE值(推荐评价标准)的最近邻居来预测用户未评分项目的评分,并且通过实验验证了这种方法比单纯用基于用户的协同推荐算法具有更高的推荐质量。在推荐阶段本文采用了多模型推荐方法。另外,系统采用众数法解决推荐系统中冷启动(新项目和新用户)问题,提高了推荐系统的推荐质量。仿真实验表明:改进的协同过滤推荐算法比传统的协同过滤推荐算法具有更好的推荐效果。最后利用改进的协同过滤算法设计实现了一个简单的电影推荐系统,达到了预期的推荐效果。更多还原

【Abstract】 Recently, Recommendation Systems for E-Commerce have obtained prevalent success. Collaborative filtering is one of the widely recommendation technologies of individuation._Whereas, there are some problems in Recommendation Systems for E-Commerce, such as sparse data of users’ rating, bad expansibility and lack of multiformity, can not cover all the interests of users. Meanwhile, quantity of valid information on net and sorts of zoomed commodity bring with austere challenge.This thesis has researched recommendation system of individuation and main recommendation technology of individuation, especially collaborative filtering technology which includes collaborative filtering based on users and collaborative filtering based on items. The work and innovation of this thesis has four parts as follows:Firstly, in computing user’s similarity phase, this thesis puts forward a mended similarity compute method based on users’ drifting interests which considering the impact of user’s rating time. Sequentially, it obtains the most valid nearest neighbor of active user.Secondly, reaching valid neighbor is very important. This thesis adopts users’ feature to cluster users. Then through the mended method of user’s similarity from cluster which active user in, it can reach the nearest neighbor.Thirdly, in forecast phase, the thesis adopts the nearest neighbor who can attain the lowest MAE using mended algorithm. And this new method is validated that it can more enhance accurate of commendation than collaborative filtering based on user.Fourthly, the thesis adopts several models recommendation. At the same time, it adopts mode method to solve cold start (new item or new user) so that enhancing the quality of recommendation system.Subsequently, this thesis takes an emulate experiment, analyzes results of experiment, compares the mended recommendation algorithm and tradition recommendation algorithm and proves mended recommendation algorithm is more precise than tradition recommendation algorithm.At last, this thesis designs and implements a simple recommendation system of film which uses the mended algorithm.更多还原