【作者】 张光明；

【导师】 郝润芳； 曲根峰；

【作者基本信息】 太原理工大学 ， 电子与通信工程（专业学位）， 2020， 硕士

【摘要】 由于互联网和电子商务的迅猛发展,物品信息量正以惊人的速度增长,信息过载问题严重地影响着人们的日常生活。如何能够快速便捷地为用户提供想要的物品已经成为这个时代热点解决问题之一。推荐系统可以克服信息过载,帮助用户提供选择。经过相关研究人员的不断研究与应用,推荐系统得到了长足的发展。21世纪随着电子科学技术的更新、发展,数据量已经进入一个新的阶段,整体数据呈现一个产生时间急、传播速度快、更新周期短的形式。传统的推荐技术已经不能满足当下的需求,因而暴露出较多问题,其中较为典型的问题有冷启动,数据稀疏,缺乏推荐的多样性等,这些问题均严重影响了推荐结果的质量。针对上述问题,本文主要工作分为如下三个部分:(1)概述了本课题选题的背景、意义以及该研究方向的发展现状;具体阐述了推荐系统的基本定义,个性化推荐系统的工作流程;具体介绍了几种经典推荐算法,和两种社会化推荐算法模型;分析了不同推荐算法评价指标和它们的应用场景。(2)提出一种基于用户评分偏好的社会化推荐算法模型。首先,本文依据用户个人的评分偏好为置信度设置动态度量方法,不仅考虑了用户给出评分的置信程度,而且还考虑了不同用户评分偏好。其次,拓展社会信任关系,从中推导出间接相关朋友,即隐语义朋友,来提升社会信任信息在推荐算法中的应用。本文将用户评分矩阵映射为用户-物品二值网络,然后通过网络中不同用户与某一物品之间的交互反馈信息,构建协同用户网络,最后从中计算出与目标用户相似的隐语义朋友,以构建相似用户矩阵来预测用户评分。通过在三个电影公开数据集上的实验,证明本节提出的算法相比其它未加用户评分偏好置信度和利用明确的信任关系推荐算法,在评分预测误差上有很大提升。(3)提出一种基于知识图谱的社会化推荐算法模型。将用户-物品交互网络与用户信任网络相结合,构建成异构知识图谱。首先,为随机游走过程设计元路径,收集正、负语料库。其次,从收集到的语料库中推导出与目标用户相关的正隐语义朋友和负隐语义朋友。最终,使得计算得到的隐语义朋友的偏好与目标用户的偏好更加相似。同样在三个公开电影数据集上对提出的算法模型进行验证,相比上节提出的推荐算法,以及其它基于明确信任关系的推荐算法,在评分预测准确度上又有进一步的提升。更多还原

【Abstract】 With the rapid development of Internet and e-commerce,the amount of information is increasing at an alarming speed,and then,overloaded information has seriously affected people’s daily life.How to provide users the desired items quickly and conveniently has become one of the hot issues in this era.The recommendation system can overcome the problem of overloaded information and provide users what they want.Through continuously research and applications by relevant researchers,the recommendation system has made great progress.However,with the development of electronic science and technology in the 21 st century,the amount of data has entered a new stage,in which generation time and update cycle are getting shorter and transmission speed is getting faster.The traditional recommendation technology has been unable to meet the current needs.And also many problems have been exposed,especially,cold start,sparse data,and lack of recommendation diversity,and they seriously affect the quality of recommendation results.In view of the above problems,the main work of this thesis is as follows:(1)The background,significance and development status of this research are summarized;The basic definition of recommendation system and the workflow of personalized recommendation system are expounded.Several classical recommendation algorithms and two social recommendation algorithm models are introduced.The evaluation indexes of different recommendation algorithms and their application scenarios are analyzed.(2)A social recommendation algorithm model,which is based on user rating preference is proposed.Firstly,under the dynamic measurement method,this thesis sets about the confidence degree by using the users’ personal rating preference,in which not only the degree of confidence given by users,but also the rating preference of different users are considered.Secondly,the social trust relationship is expanded,from which the indirectly related friends,namely the cryptic friends,are derived to improve the application of social trust information in the recommendation algorithm.In this thesis,the user rating matrix is mapped to the user-item binary network,and then the collaborative user network is constructed by the interaction feedback information between different users and an item in the network.Finally,the latent friends,who are similar to the target users,are calculated from it,and the similar user matrix is constructed to predict the user rating.Through experiments on three open movie data sets,it is proved that the algorithm proposed in this section has a great improvement in the prediction error of rating compared with other algorithms,which did not add the confidence degree of user rating preference and made use of explicit trust relationship recommendation algorithm.(3)A social recommendation algorithm model,which is based on the knowledge map is proposed.The heterogeneous knowledge map is constructed by combining the user-item interaction network with the user trust network.The meta-path is designed for the random walk process,and the positive and negative corpus are collected to deduce the positive and negative latent friends,who are related to the target users,so that the calculated cryptic friends’ preferences are more similar to those of the target users.The proposed algorithm model was also validated on three movie public data sets.Compared with the recommendation algorithm proposed in the previous section and other recommendation algorithms based on explicit trust relationship,the accuracy of score prediction was further improved.更多还原