【作者】 钱付兰；

【导师】 张燕平；

【作者基本信息】 安徽大学 ， 计算机科学与技术， 2016， 博士

【摘要】 随着信息技术和互联网的飞速发展,网络资源变得极大丰富,多渠道的信息来源使得信息的产生变得更加便捷。随之带来的海量信息在给人们提供丰富资讯的同时,过载信息(overload information)也带来人们选择的困难。推荐系统作为一种能够解决信息过载问题并可提供个性化服务的有效工具近年来得到了广泛地关注。协同过滤是目前运用最为成功的推荐技术,近年来在此之上的新算法层出不穷,随着网络规模的扩大,用户和商品人数的激增和社交媒体的加入,协同过滤方法面临着诸多挑战。商空间粒度计算理论利用人类处理复杂问题的求解思想,将问题高度抽象成论域,属性和结构三元组(X,f,T)进行描述,其采用由粗到细粒化逐步求解的分析方法,对复杂问题的求解具有重要的意义。本文利用商空间模型对推荐问题进行表示,通过粒化可改善推荐性能。首先依据商空间理论对推荐问题进行总体描述和说明,接着分别依据属性和结构对协同过滤算法进行研究,重点研究了推荐系统中的数据稀疏性问题、实时性问题、鲁棒性问题、社交关系对推荐系统影响等若干问题,给出了商空间模型的对应描述,从粒计算的角度对这类问题提出了相应的解决方法。本文的主要研究工作概括如下：(1)给出了商空间理论对于推荐系统问题的整体描述；利用商空间中三元组(X,f,T)表示问题的论域、属性和结构。将推荐问题看成用户评分为论域X,用户相关特征为属性f,用户与用户相互关系为结构T的系统,构建了用户为核心的商空间模型。并进一步给出了常见的协同过滤推荐算法的粒度视角下的描述。以该该模型作为基础,对以用户声誉、用户间评分关系、用户间社交关系等作为属性和结构,为论文提及的推荐系统中若干具体问题研究与求解奠定了基础。(2)研究了用户声誉属性对多种推荐系统的影响；隐语义分解模型以商空间的视角可以看成在某一最优准则下生成两个较小粒子,并进一步合成较大粒子的过程,我们将用户声誉属性代入到最优准则中,将其用于不同的隐语义分解模型中,考察用户声誉值对不同推荐系统的影响。本文利用历史评分迭代学习得到用户声誉值,并将其引入到一般推荐系统和社交推荐系统中,提出了结合声誉的LFM模型(LFMrep)和基于声誉的社交推荐PMF模型(SoRS)。实验结果表明使用用户声誉可以去除系统由于用户评分不严谨带来的噪声,提高系统推荐精度；在一般推荐系统中对声誉值过高的用户进行抑制可以提高系统的鲁棒性,增强系统的抗攻击能力；在社交推荐中引入声誉系统在数据稀疏训练集不足的情况下推荐精度亦可得到保证。另外使用粒化思想,提出了基于用户声誉粒化的社交推荐算法(SrBug),实验结果表明在短推荐列表推荐精度无明显变化的情况下,推荐响应优势明显。(3)研究了利用粒度思想对推荐问题中用户邻域进行结构粒化的方法；本文将用户之间的相似性关系视为推荐问题中具有无权特性的网络结构关系,使用商空间粒化的思想构造出用户邻域,并在用户邻域中进行推荐。提出了一种依据无结构的评分数据进行建模,映射出用户问评分关系网络,并使用社团发现算法对其用户邻域进行结构粒化的方法(CUCRA),实验结果表明,利用结构粒化的方法,可以更有效发现用户近邻,在不降低精度的前提下,算法具有较好的线上响应时间。进一步地我们使用分层聚类的思想对用户邻域进行优化,形成层次粒化用户邻域模型(HGUCRA),相关实验表明该模型中的推荐结果在获得较优时间响应的同时也获得了较好的推荐精度。(4)以模糊等价关系定义了社交网络中的“三度”之内的社交关系及其分层递阶结构,深度挖掘了三度之内的社交关系对推荐系统精度的影响。本文选取用户间社交关系及其相似性作为推荐问题中的有权特性的网络结构进行研究,依据模糊等价关系的定义得到以步长为截集的分层递阶结构,提出了基于上下文子图传播的隐含关系的社交推荐算法(IRSubNet),该算法将“三度”之内的隐含关系和明确关系整合到了一个统一的框架中来,相比其他推荐算法,IRSubNet深度挖掘了三度之内的社交关系,定义了社交网络中任意两点基于上下文子图的社交相似性的计算方法,重点分析了隐含关系在推荐系统中对精度的影响。在两个真实的具有社交网络的数据集上所做的实验结果表明,该算法的精度有明显提升。更多还原

【Abstract】 With the rapid development of the information and network technology, the resources become more and more abundant. The information sources of multi channels make the information more convenient. The massive information is available to provide rich information, while the overload information have made people too difficult to choose the useful information.As one of the effective tool that can solve the information overload problem and provide the personalized service, recommender systems have received extensive attention in recent years. Collaborative filtering is one of the most widely used and successful methods for recommendation. Based on the collaborative filtering, the new algorithms are emerging in an endless stream. With the expansion of network scale, increase sharply in number of users/items and use of social media, the collaborative filtering faces several opportunities and challenges.Utilizing the though of human deal with complex problems, Quotient Space Theory highly abstract the problems into a triple (X,f, T). This theory utilizes the analysis method which is from coarse to fine to solve the problems step by step and has the great significance to the solution of complex problems.In this dissertation, the recommendation is expressed by quotient space triple.And it can improve the recommendation accuracy through granulation. Firstly, we give the general description and explanation of recommendation problem according to the quotient space theory. Then we make a research into the collaborative filtering partly according to the attribution, structure etc., and mainly researched the problems such as Data sparsity, real-time, robustness and the effect of social relation in recommender system. Given the corresponding description of quotient space model, we proposed the solving methods against the problems mentioned above from the granular computing perspective.The dissertation includes:1. Give the unitary definition of recommender systems based on quotient space theory.Utilizing the a triple (X,f,T) in quotient space theory, the recommender system can be considered as the system that the user ratings, the user relevant characters and the relations between users are respectively viewed as the domain: X, the attribute function: f, andthe structure of the elements in the domain:T. Thus the new model of recommender system based on the quotient space model with a core of usercan be built. Furthermore the model description of the basic collaborative filtering algorithms in the view of granular aregiven in the dissertation. Based on this model, we research and solve some concrete problems taking user reputation, the rating relations between users, and the social relations between users as attribution and structure.2. Research the effect of user reputation on various recommender systemsLatent factor model can be considered as the produce of generating two relatively smaller granules under a certain optimal rules and further synthesizing larger granule from the perspective of quotient space. We introduce the user reputation into the optimal rules which be used in different latent factor model, and investigate the effect of user reputation on various recommender systems.We proposed LFMrep that is the LFM model combined user reputation and SoRS that is PMF model based user reputation. And the user reputation can be obtained by iteratively learning from historical ratings, which has been introduced into normal recommender systems and social recommender systems as mentioned above in the dissertation. Experimental results show that our methods have improved in terms of prediction accuracy criteria by using the user reputation can effectively remove noise because of user less rigorous ratings in recommender systems. In the normal recommender system, it can improve the robustness and vulnerability to attacks of the system by weakening the attribution of high reputation user. And in social recommender system, under the insufficiency of training set, the recommendation accuracy can be maintained by introducing the user reputation.. In addition, we also proposed SrBug (Social Recommendation Based on User Reputation Granulation) which has advantage in online response with short recommendation list under no obvious change of accuracy.3. Give the method of user domain structure granulation in recommendation using the granularity idea.The similarity between users is viewed as the network structure relation with no weight character in recommendation. Using the granulation idea in quotient space theory, the user domain can be structured where the recommendation generated in. We proposed the CUCRA (Community-Based User Domain Model Collaborative Recommendation Algorithm), through modeling by ratings with no structure, mapping the rating relation network between users, and granulating the user domain by using community detection algorithm. Empirical results demonstrate that it can find user domain neighbor more efficiently by using the method of structural granulation. And it also has quick online response without losing accuracy. Furthermore we proposed the HGUCRA (Hierarchical Granulation-based User Domain Model Collaborative Recommendation) in order to optimize the user domain by hierarchical clustering methods. Related experiments show that the proposed model can obtain a better recommendation accuracy, on the meantime it also has the better online response.4. Define the social relation within "three degrees" and hierarchical structure by fuzzy equivalency relations, deeply mine the effect of social relation within "three degrees" on the recommendation accuracy.Selecting the social relation and similarity between users as the network structure with weight character in recommendation, we proposed the IRSubNet which incorporate the implicit relations based on context-aware subnet-work propagation for SoReg. According to the definition of fuzzy equivalency relations, it can be obtained hierarchical structure with step in network as cut set. The proposed method integrated implicit relations and explicit relations within "three degrees" into unified framework. In comparison with other algorithms, IRSubNet deeply mine the effect of social relation within "three degrees", define the social similarity between any two nodes in social network based on context-aware subnet-work, and emphasizes the effect of implicit relation on the recommendation accuracy. The experimental results shows that IRSubNet can obviously improve the recommendation accuracy on the two real datasets.更多还原