【作者】 刘坤；

【导师】 陈晓云；

【作者基本信息】 兰州大学 ， 软件工程·软件工程， 2017， 硕士

【摘要】 在数据时代,人们生活在一个信息过载的互联网空间中,面对海量的、重复的、有用的、无用的信息,人们很难从中选择自己感兴趣的东西。为了解决这个问题,学者们提出了推荐系统理论。推荐系统通过分析用户的兴趣爱好和需求,将用户感兴趣的产品、信息或者服务推荐给用户,从而有效地缓解了信息过载问题带来的影响。在推荐系统的研究历程中,协同过滤推荐、基于内容的推荐、基于知识的推荐和混合推荐等理论被相继提出。此外,为了更加有效地解决信息过载问题,学者们还在这些推荐理论的基础上不断创新,将数据挖掘、机器学习、深度学习等方法运用到推荐系统中。推荐系统的研究面临着诸多挑战,其中数据稀疏性、冷启动、大数据处理与增量计算、系统脆弱性是其面临的主要问题。冷启动问题分为非纯冷启动问题和纯冷启动问题,即刚加入系统的用户有较少的产品评价记录甚至没有。面对这些新用户,推荐系统可利用的他们的产品评价信息非常少或是没有,因此准确地给他们推荐产品变得极其困难。由于冷启动问题的不可避免性及其对推荐准确率造成的影响,为此,本文提出了两个新的推荐模型:NPBM和NDBM,分别用于解决非纯冷启动问题和纯冷启动问题。对于非纯冷启动问题,因为冷用户只有较少的产品评价记录,推荐系统很难准确地分析出他们的产品偏好,因此本文提出了结合机器学习方法的NPBM推荐模型用于降低该问题带来的影响。NPBM模型根据用户的产品评价信息准确地找到冷用户的最近邻,然后将这些最近邻用户喜欢的产品推荐给冷用户。对于纯冷启动问题,因为冷用户没有任何产品评价记录,推荐系统无法分析冷用户的产品偏好。为此,本文提出了基于人口统计学理论的NDBM推荐模型用于解决该问题。NDBM模型解决纯冷启动问题的思路是利用用户的人口统计学特征信息准确地找到冷用户的相似用户,然后将这些相似用户喜欢的产品推荐给冷用户。本文在真实的数据集上进行了模拟推荐实验,实验结果显示NPBM和NDBM模型面对冷启动问题仍然具有较高的推荐准确率,它们能有效地缓解冷启动问题带来的影响。更多还原

【Abstract】 In the data age,information in the cyberspace is overloaded.In the face of massive,repetitive,useful,useless information,it is difficult for users to choose the products that they are interested in.In order to solve this problem,scholars have put forward recommender system.Recommender system can alleviate the impact of the information overload problem by analyzing users’ hobbies and needs,and recommending products,information or services that users are interested in to them.In the research process of recommender system,the theories of Collaborative Filtering algorithm,Content-Based recommendation,KnowledgeBased recommendation and Hybrid recommendation has been put forward.In addition,in order to more effectively solve the problem of information overload,continuous innovations of recommender theory have been proposed.What’s more,technologies in data mining,machine learning,deep learning and other methods have been applied to recommender system.There are many challenges in the research of recommender system,such as data sparseness,the cold-start problem,large data processing in incremental calculation,and system fragility.The cold-start problem can be divided into the non-pure cold-start problem and the pure cold-start problem,that is,users that have just joined the system have less product evaluation records even no records.For these new users,it is extremely difficult to recommend products that are available to them.Due to the unavoidability of the cold-start problem and its influence on recommendation accuracy,in this thesis,the NPBM recommender model and the NDBM recommender model have been proposed to solve the problems of the non-pure cold-start and the pure cold-start,respectively.For the non-pure cold-start problem,recommender system is difficult to accurately analyze the product preferences of cold users,because cold users have only a few product evaluation records.This thesis proposes the NPBM recommendation model which based on the machine learning method,to solve the problem.The idea of the NPBM recommender model is to accurately locate the nearest neighbors of the cold user based on the users’ product evaluation records,and then recommend the products that his nearest neighbors prefer to the cold user.For the problem of pure cold-start,recommender system can not analyze cold users’ product preferences because they do not have any product evaluation records.To this end,this thesis proposes the NDBM recommender model which is based on the demographic theory to solve this problem.The NDBM recommender model uses the demographic characteristics to accurately find the similar users of cold users,and then recommend the products that these similar users like to the cold users.In this thesis,a large number of simulation experiments are carried out on real datasets.The experimental results show that the two recommender models have high recommendation accuracy when facing the cold-start problem.That is to say,our models can effectively alleviate the influence of the cold-start problem.更多还原