基于统计推理的复杂网络社区结构分析

发布时间：2018-02-21 12:43

本文关键词： 复杂网络社区发现统计推理概率模型非参贝叶斯好友推荐　出处：《哈尔滨工业大学》2016年博士论文　论文类型：学位论文

【摘要】：现实世界中的许多复杂系统,如互联网、物联网、食物链网、神经网络以及社会网络等,都可以描述成由节点集通过边连接构成的复杂网络。社区发现作为复杂网络的一项重要研究内容吸引了多个领域的研究人员从不同视角对其进行深入研究。尽管近年来涌现出一大批新颖的社区发现方法,但依然存在许多问题,包括网络的重叠社区发现、异质网络的社区发现、网络社区的自动探索以及社区发现的应用等。本文提出一系列基于统计推理的方法来进一步研究解决上述难题,研究内容主要包括重叠社区发现和社区自动探索两个方面。重叠社区发现主要针对符号网络和加权网络,其中符号网络考虑了边的正负极性,加权网络考虑了边的强弱程度;社区自动探索主要针对同质网络和异质网络,其中同质网络由单一类型节点和单一类型边组成,异质网络由多类型节点(如带属性节点)或多类型边(如多维度边)组成。具体内容包括:第一,研究了基于混合模型的符号网络重叠社区发现。现有的符号网络社区发现方法都仅允许网络的节点属于一个社区,但真实世界中网络的节点可能同时属于多个社区(即重叠社区)。针对这个问题,本文提出了一种符号概率混合模型用于发现符号网络的重叠社区。它用混合模型同时描述正边和负边的生成过程,其中正边存在于社区内,负边存在于社区间。该模型不仅能发现符号网络的重叠节点,还能提供节点属于社区的隶属度信息。在大量的符号网络上的实验表明该模型能发现符号网络的重叠社区并取得较优的性能。第二,研究了基于贝叶斯的加权网络重叠社区发现。现有的混合模型方法能比较好地发现加权网络的重叠社区,但在某些网络上会出现一些节点不属于任何社区,从而导致模型不能适用于网络。本文提出了一种贝叶斯混合网络模型用于加权网络的重叠社区发现。该模型通过为模型参数引入先验分布解决了上述问题。在人工合成网络和真实网络上的实验表明该模型能发现加权网络的重叠社区并取得较优的社区划分效果。第三,研究了基于非参贝叶斯的网络社区自动探索。现有的社区发现方法要么需要预先假定网络的结构类型(如社团结构、多部图结构或混合结构等),要么需要预先指定网络的社区数目,但真实世界中许多网络的结构类型和社区数目都是未知的。针对这个问题,本文提出了一种贝叶斯非参混合模型用于自动探索网络的社区结构类型和社区数目。它将具有探索网络社区结构类型能力但需要预先指定社区数目的混合模型拓展到非参框架下,利用狄利克雷过程自动确定其社区数目。在大量的人工合成网络和真实网络上的实验表明该模型能自动探索网络的结构类型和社区数目并取得较优的社区划分效果。另外,基于该模型的好友推荐系统取得了很好的效果。第四,研究了基于非参贝叶斯的异质网络社区自动探索。本文分别从带属性节点网络和多维度网络两方面研究了异质网络的社区自动探索。带属性节点网络社区自动探索的一个主要难点在于如何有效利用节点的属性信息来提高探索效果。本文提出了一种贝叶斯非参属性模型用于自动探索带属性节点网络的社区结构类型和社区数目。它通过共享变量的方式同时融合了节点的连接和属性信息。在真实和人工合成的带属性节点网络上的实验表明该模型能自动探索网络的结构类型和社区数目并取得较优的社区划分效果。现有的多维度网络社区发现方法的主要不足在于需要预先假定网络的结构类型(如社团结构等)。本文提出了一种多维度贝叶斯非参混合模型用于自动探索多维度网络的社区结构类型和社区数目。该模型先利用网络社区自动探索方法从每一维度抽取出结构特征信息,然后利用现有的聚类方法对融合后的结构特征进行聚类得到多维度网络的结构类型和社区数目。在真实和人工合成的多维度网络上的实验表明该模型能自动探索网络的结构类型和社区数目并取得较优的社区划分效果。
[Abstract]:Many complex systems in the real world, such as the Internet, the Internet of things, the food chain network, neural network and social network, complex network can be described by sets of nodes connected through edges. The community found as an important research content of complex networks has attracted researchers in various fields from different perspectives in-depth study on it. Although in recent years the emergence of a large number of methods to find novel community, but there are still many problems, including the overlapping community network, heterogeneous network community, network community to explore and automatic community discovery applications. This paper proposed a series of methods of statistical inference based on the further study to solve the problem the main research contents, including two aspects of automatic discovery and community overlapping community. Overlapping community discovery mainly for symbolic network and weighted network, the symbol Considering the network side of the polarity of the weighted network considering the degree of the edge of the community; automatic exploration mainly for homogeneous network and heterogeneous network, wherein the network from a single type of homogeneous nodes and single side, heterogeneous network consists of many types of nodes (such as with an attribute node) or multiple types of edges (such as multi dimension side). The specific contents include: first, the research found that symbolic network hybrid model based on symbolic network community. Overlapping community existing node discovery methods are only allowed to belong to a community network, but the network in the real world, the node may also belong to multiple communities (i.e. overlapping communities). Aiming at this problem, this paper puts forward a symbol probability mixture model to detect overlapping communities. It uses symbolic network hybrid model and description generation process and the negative side of the positive side, the positive edge exists in the community, the negative side exists Between communities. This model can not only find the overlapping node symbol of the network, but also provide the membership information of nodes belonging to the community. In the symbolic network on a large amount of experiments show that the model can detect overlapping community symbolic network and obtained the performance better. Second, the research found that the weighted network based on Bayesian overlapping communities. Hybrid model of existing methods can better detect the overlapping community weighted network, but in some network there will be some nodes do not belong to any community, resulting in the model can not be applied to the network. This paper proposes a Bayesian network model for hybrid overlapping community discovery. The weighted network model by introducing a priori model parameters to solve the above problems. The distribution in synthetic and real networks. Experimental results show that the model can detect overlapping communities and obtain the optimal weighted network community Results. Third, the non parametric Bias community network. Automatic exploration based on the existing community discovery method or structure types presuppose the network (such as community structure, multi partite graphs or mixed structures etc.), or need to specify the number of network community, but the community structure and the number of many real networks the world is unknown. To solve this problem, proposed mixed model is used to automatically explore the network number of community structure and community type non Bias. It will have a mixed model but need to explore a pre specified number is extended to the non parametric community under the framework of network community structure, automatically determine the number of communities by using the Dirichlet process. In the network a large number of synthetic and real network experiments show that the model can automatically explore the structure of the network and the number of community types And get a better community division. In addition, the model of friend recommendation system achieved good results. Based on the fourth, on the heterogeneous network community non parametric automatic exploration based on Bias. This paper respectively from the attribute node network and multi dimension network in two aspects of heterogeneous network community. An automatic exploration a major difficulty with an attribute node network community is to automatically explore how to effectively use the attribute information of the node to improve the exploration effect. This paper proposes parameter attribute model for automatic exploration with an attribute node of the network structure and the number of community community than a Bias. It's way through shared variables while fusion node attribute information in real and synthetic attribute nodes on the network experiments show that the model can automatically explore the structure type and the number of community network and make Community division effect and existing multi dimension network community detection method is the main problems need to presuppose the network structure (such as community structure and so on). This paper proposes a multi-dimensional Bayesian nonparametric mixture model for automatic multi dimension exploration network community structure and community type number. The first use of network model to explore the method of automatic extraction of community structure information from each dimension, and then use the existing clustering methods of structure types and the number of community clustering multi dimension network structure of the fused features. In multi dimension network on real and synthetic experiments show that the model can automatically explore the structure type and the number of network community and get better. Community Division

【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：O157.5

【相似文献】