【作者】 王聪；

【导师】 孙静宇；

【作者基本信息】 太原理工大学 ， 计算机科学与技术， 2018， 硕士

【摘要】 随着互联网的快速发展,网上信息量越来越多,能为用户带来精准推荐的系统可以大大提升用户体验,个性化推荐算法成为了研究的热点问题。协同过滤是应用最广泛的推荐算法,它根据用户的历史行为数据对用户的兴趣做分析,进行相应推荐,但其研究中存在严重的数据稀疏问题,并且现有的算法对推荐的实时性考虑不足,这都是目前急需解决的课题。本文对当前的推荐算法进行研究,分析了当前推荐系统的研究现状,同时对深度学习在推荐系统领域的研究做了深入分析,意在将深度学习中栈式降噪自动编码器在自动特征提取方面的能力应用到推荐系统中,来缓解推荐系统存在的数据稀疏问题,同时提高推荐系统实时推荐的性能。为实现这一目的,本文做了多方面研究:(1)基本理论研究首先介绍了几种典型的推荐算法,并着重从协同过滤推荐算法入手,详细介绍了基于用户的协同过滤推荐算法和基于项目的协同过滤推荐算法。其次介绍了几个典型的深度学习框架,然后列举了几个典型的基于深度学习的推荐系统。(2)增量均模型及其改进研究通过分析本课题组前期提出的均模型及其增量扩展研究,发现在进行相似度计算时未考虑用户兴趣漂移问题,同时,在利用传统kNN方法选取近邻对象集时存在着一定的局限性,针对这两个问题,本文提出了一种动态选择的时间均模型。该模型引入近邻因子和时间函数,采用动态选择的方法选取项目的近邻集合,利用时间函数实现时间加权下的增量均模型改进。(3)时间均模型及其改进研究深度学习具有强大的特征提取能力,本文详细阐述了栈式降噪自编码器(Stacked Denoising Auto-Encoder,SDAE)的工作原理,并提出将其用于协同过滤推荐中学习用户(项目)的深层特征表示,但SDAE不能很好地利用用户对项目评分的时间属性。在(2)中提出的时间均模型,具有很好的处理大数据的能力以及实时推荐性能,但在特征提取过程中,由于数据稀疏问题的存在,会发生信息失真的情况。针对以上问题,提出了一种基于SDAE和时间均模型混合的协同过滤推荐算法,充分利用SDAE的自动特征学习能力以及时间均模型在大数据下的快速计算能力,较好地弥补了两者各自的不足之处。(4)均模型改进研究的实现本文在3个标准数据集上分别进行实验,通过对比传统方法和本文提出的方法来验证a)时间属性在实时推荐中的重要性;b)利于动态的方法选择用户(项目)近邻集合的必要性;c)将深度学习模型SDAE自动提取用户(项目)深层特征方面的优势应用到推荐系统中对推荐性能的影响。实验结果表明,本文提出的两种均模型改进方法进一步提高了推荐系统的推荐性能。更多还原

【Abstract】 With the rapid development of the Internet,the amount of information on the Internet is more and more.The system that can bring accurate recommendation for users can greatly improve the user experience.Personalized recommendation algorithm has become a hot topic of research.Collaborative filtering is the most widely used recommendation algorithm.It analyzes and recommends users’ interest according to the user’s historical behavior data,but there is a serious problem of data sparsity in its research,and the existing algorithms are not enough to consider the real time of the recommendation,which are all the problems to be solved urgently.In this paper,the current recommendation algorithm is studied,and the current research status of the recommendation system is analyzed.At the same time,the application of deep learning in the field of recommender system is deeply analyzed.It is intended to apply the capability of the automatic feature extraction of stacked denoising auto-encoder in deep learning to the recommendation system to alleviate the problem of data sparsity,meanwhile,it can improve the real-time recommendation performance of recommendation system.To achieve this goal,there are many researches done in this paper.(1)Basic theory researchThis paper first introduces several typical recommendation algorithms,and focuses on collaborative filtering recommendation algorithm,and introduces user based collaborative filtering recommendation algorithm and project based collaborative filtering recommendation algorithm in detail.Secondly,severaltypical deep learning frameworks are introduced,and several typical recommendation systems based on deep learning are listed.(2)Incremental mean model and its improvementThrough the analysis of the mean model and its incremental expansion proposed by our group,we find that the problem of user interest drift is not considered when the similarity calculation is carried out.At the same time,there are some limitations when using the traditional kNN method to select the nearest neighbor object set.In this paper,a dynamic selection time mean model is proposed in this paper.The model introduces the nearest neighbor factor and the time function,uses the dynamic selection method to select the nearest neighbor set of the project,and uses the time function to improve the incremental mean model under the time weighting.(3)Time mean model and its improvementDeep learning has powerful feature extraction ability.This paper describes the working principle of Stacked Denoising Auto-Encoder(SDAE)in detail.And it is used to learn the deep feature representation of users(projects)in collaborative filtering recommendation.But SDAE can’t make good use of the rating’s time property.The time mean model proposed in(2)has a good ability to deal with large data and real-time recommendation performance.However,in the process of compression vector,information distortion will occur because of the existence of data sparsity.In view of the above problems,a hybrid recommendation algorithm based on SDAE and time mean model is proposed,which makes full use of the automatic feature learning ability of SDAE and the fast calculation ability of time mean model in large data.It makes up for the shortcomings of both of the two.(4)Implementation of the improved mean modelThis paper carries out experiments on 3 standard data sets,verifies a)the importance of time attributes in real-time recommendation;b)the necessity ofdynamic methods when select the nearest neighbor set of users(projects);c)the capability of the automatic feature extraction of SDAE by comparing traditional methods and the methods proposed in this paper.The experimental results show that the two improved models proposed in this paper further improve the recommendation performance of the recommender system.更多还原