极性异构信息网络的联系预测技术研究

发布时间：2018-04-21 18:32

本文选题：链接预测 + 极性预测　；参考：《山东大学》2014年硕士论文

【摘要】：我们生活在一个相互关联的世界。大多数数据或信息对象、组件等是内部关联或者相互作用的,形成了无数的、巨大的、相互关联的复杂网络。不失一般性,相互关联的网络称为信息网络。分析和挖掘信息网络已经成为计算机科学、社会学、生物学等领域的研究人员广泛关注的课题。信息网络分为同构信息网络和异构信息网络。同构信息网络只有一种类型的节点和一种类型的关系,如在朋友关系网络中,节点都是人这一类型,边只表示朋友关系。然而,现实中的大部分网络都是异构的。在异构信息网络中,节点有多种类型,而不同类型的节点之间的关系属于不同的类型,如IMDB网络中,有电影、导演、演员等不同类型的节点和电影-导演之间的执导关系、电影与演员之间的参演关系等具有不同语义的关系类型。随着网络的发展,人们在网络社交时越来越多地表达自己的情感,因此网络中的边便有了极性,即边是正的(表示信任、喜欢、朋友等关系)或负的(表示不信任、不喜欢、反对等)。我们称有极性的异构信息网络为极性异构信息网络。信息网络已有了很多的分析和挖掘方法的研究,联系预测是其中的一个重要任务。在极性异构信息网络中,联系预测包含链接预测和极性预测,分别预测边的存在性和极性。链接预测在分析演化网络、推荐、聚类等领域有重要的价值,极性预测可以应用在推荐、决策制定、网络演化模型等众多领域。虽然链接预测和极性预测都有了很多的研究成果,但大多数链接预测都以非极性信息网络为基础,极性预测多以同构信息网络为基础,而现实中大多数网络是极性异构信息网络,所以如何解决极性异构信息网络中的联系预测问题成为新的挑战。本文针对极性异构信息网络,探索了该网络下的联系预测问题,主要工作可归结于以下几点： 1.提出了极性异构信息网络的链接预测解决方法。在本文中,我们提出基于规则的方法,称为RulePredict来解决链接预测问题。在RulePredict模型中,我们首先系统抽取特征,特征包括促进链接存在的正特征和减弱链接存在可能性的负特征。链接是否出现服从概率为p的二项分布,p为所有特征值的函数。然后,使用基于广义最小二乘法的监督学习方法学习不同特征对应的权重。将学习到的权重应用到测试数据中来预测链接是否存在。 2.提出了极性异构信息网络的极性预测解决方法。我们提出一个新的方法HeteSign来解决极性预测问题。首先定义不同关系下的节点相似值,每个节点相似值看作一个特征,有相对应的权重。节点间的相似度定义为特征和权重的数学表达式。计算链接的极性得分,根据得分判断链接是正是负。得分表示为节点相似度和现有网络的链接的函数,现有的链接由于正负边的重要性不同,赋予相对应的系数。采用监督学习框架,使用极大似然估计算法求得权重和系数。 3.在真实的数据集IMDB和Epinions网络上验证上述两种方法的有效性,实验结果证明我们的方法在准确性上比其它方法表现更好。
[Abstract]:We live in an interconnected world . Most data or information objects , components , etc . are internal links or interactions , forming numerous , huge , interrelated complex networks . Unlost generality , interconnected networks are known as information networks . Analytical and mining information networks have become a subject of extensive concern to researchers in the fields of computer science , sociology , biology , etc .

The information network is divided into a homogeneous information network and a heterogeneous information network . The homogeneous information network has only one kind of node and one kind of relation , such as in a friend relationship network , the nodes are of the same type , and the relationship between the film and the actor is of different types . As the network grows , people express their feelings more and more in the network , so the edges in the network are positive ( representing trust , affection , friends , etc . ) or negative ( representing distrust , dislike , opposition , etc . ) . We call polar heterogeneous information networks as polar heterogeneous information networks .

The information network has a lot of research on the analysis and mining methods , and the contact prediction is an important task . In the polar heterogeneous information network , the contact prediction includes link prediction and polarity prediction , and the existence and polarity of the edges are predicted respectively . The link prediction has important value in the fields of analysis and evolution network , recommendation , clustering and so on . The polarity prediction can be applied in many fields such as recommendation , decision making , network evolution model and so on .

Although the link prediction and the polarity prediction have many research results , most of the link prediction is based on the non - polar information network , the polarity prediction is based on the homogeneous information network , and most networks in the reality are polar heterogeneous information networks , so how to solve the problem of contact prediction in the polar heterogeneous information network becomes a new challenge . In this paper , the connection prediction problem under the network is explored for the polar heterogeneous information network , and the main work can be attributed to the following points :

1 . In this paper , we propose a method of link prediction for polar heterogeneous information networks . In this paper , we propose a rule - based approach , called Rulemaking , to solve the link prediction problem . In the Rulemaking model , we first extract features , which include the positive features that promote the existence of links and the negative features of the possibility of weakening links . Whether the links appear binomial distributions with probability p and p is a function of all the eigenvalues . Then , the weights that correspond to different features are learned using supervised learning methods based on generalized least squares . The weights learned are applied to test data to predict whether links exist .

2 . A new method for predicting the polarity of polar heterogeneous information networks is proposed . We propose a new method HeteSign to solve the problem of polarity prediction . Firstly , we define the nodes similarity values under different relations . Each node ' s similarity value is regarded as a feature with corresponding weights . The similarity between nodes is defined as a function of the characteristics and weights . The scores are expressed as functions of nodes similarity and links of existing networks . The existing links are given corresponding coefficients due to the importance of positive and negative edges . Using the supervised learning framework , the weights and coefficients are obtained using the maximum likelihood estimation algorithm .

3 . The validity of the two methods is verified on the real data set IMDB and the Ephedra network , and the experimental results show that our method is better in accuracy than in other methods .

【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.02

【参考文献】