【作者】 李国成；

【导师】 李全龙；

【作者基本信息】 哈尔滨工业大学 ， 计算机技术（专业学位）， 2019， 硕士

【摘要】 在使用在线教育平台学习过程中,大量的学习资源会给学习者带来学习迷航、知识过载等问题,为了给学习者提供更好的学习服务,进行个性化学习资源的推荐是应对这两个问题的重要解决方案。近年来,推荐系统已经渗透到多种信息行业中,例如电商、资讯等,学习资源的推荐与新闻、商品推荐有许多相似之处,但又有着其独特的限制与需求。在其他推荐领域的实践证明,用户之间的社交关系可以提高推荐效果,而在线教育平台同样提供讨论区等用户交互场景,教育数据包含社交关系。为此,本文进行了信任关系辅助的在线教育资源推荐方法研究,包括:在线教育平台信任建模方法研究、基础教育资源推荐方法研究和基于信任关系的推荐方法研究。在线教育平台信任建模是本文的研究基础,支撑后续基于信任推荐方法的设计与实现。为了解决了在线教育平台信任关系提取、信任网络的构建、信任网络的扩张这三个信任建模过程中的核心问题,本文定义了由信任度、可靠度组成的二维信任意见,基于环论定义了信任意见的聚合与扩张基本运算,并利用环的性质设计实现信任建模算法并进行优化,建立可解释的、信息丰富的信任网络。进行学习资源协同过滤推荐方法的研究是因为基于信任的推荐具有局限性,只关注于社交行为导致无法充分利用资源与用户的个性化信息,例如学习兴趣、知识结构等。本文在anonymisedData、学堂在线数据集上开展基于记忆的协同过滤、对数几率回归、因子分解机三种基础方法的课程推荐实验,同时利用项目组在用户画像问题上的研究成果,将用户学习态度标签加入到输入特征向量中。通过对比准确率与召回率,最终确定因子分解机作为学习资源的基础推荐模型。基于信任建模的结果,本文设计并实现了基于协同过滤的信任推荐算法,将信任推荐与基于用户协同过滤推荐整合,并在Epinions数据集上进行实验,验证了本文提出的信任建模方法对推荐的价值,并通过对照实验,证明了混合推荐方法要优于单一推荐方法。提出了基于因子分解机的混合推荐模型,在用户向量整合了基本属性、个性化标签、信任关系、学习能力等特征,同时参考工业推荐架构,设计了混合多种方法的在线学习资源推荐系统架构。更多还原

【Abstract】 When studying on online education platform,too many learning resources may lead some problems for learners,such as learning lost,knowledge overload.In order to provide learners with better learning services,personalized learning resources recommendation is an important solution to these two problems.In recent years,recommendation system has penetrated into many information industries area,such as e-commerce,news and so on.There are many similarities between recommendation of learning resources and recommendation of news and commodities,but recommendation in education area has its own unique limits and needs.Practice in other recommendation areas has proved that social relationships among users can improve the recommendation effect,while online education platform also provides user interaction scenarios such as discussion community.Based on the above considerations,this paper will research and implement the recommendation method of online educational resources based on trust relationship assistance.Specifically,this research will be carried out in the following three methods: online education platform trust modeling method,basic education resource recommendation method and trust relationship-based recommendation method.Trust modeling of online education platform is the research foundation of this paper,which supports the design and implementation of trust-based recommendation.In order to solve three essential problems in trust modeling on online education platform: trust relationship extraction,trust network construction and trust network expansion,this paper defines two-dimensional trust opinions consisting of trust value and reliability value,and defines the basic operations of aggregation and expansion of trust opinions based on ring theory.Then we design and optimizes trust modeling algorithm utilizing the properties of ring and establish interpretable and informative trust networks.This paper also studies collaborative filtering recommendation methods of learning resources,because of the limitation of trust-based recommendation.Trust-based recommendation only focuses on social behavior,which makes it not able to make full use of personalized information of resources and users,such as learning interest,knowledge structure.In this paper,we carry out experiments on anonymised data and xuetangx datasets.There are three experiments on basic recommendation methods: memory-based collaborative filtering,logistic regression and factorization machine.When recommending by logistic regression and factorization machine,we use the other project teammates’ research results on user portraits,and add user learning attitude tags a to the input feature vector.By comparing and analyzing the accuracy and recall rates,we finally choose FM as the basic learning resources recommendation method which gets the best performanceUtilizing the results of trust modeling,this paper designs and implements a trustbaesd recommendation algorithm based on collaborative filtering which integrates trustbased recommendation with user-based collaborative filtering recommendation.We carry out experiments on Epinions dataset.The results verify the value of the proposed trust modeling method for recommendation,and prove that the hybrid recommendation method is superior to the single recommendation method.But a hybrid recommendation model based on factorization machine is finally proposed,which integrates the basic attributes,personalized labels,trust relationship,learning ability and other characteristics in user vector.What is more,we refer to the industrial recommendation architecture and design an online learning resources recommendation system,which merges plenty of recommendation methods.更多还原