无监督学习框架下学习分类器系统聚类与主干网提取方法研究

发布时间：2018-02-23 16:36

本文关键词： 机器学习无监督学习聚类分析学习分类系统集成学习复杂网络主干网提取　出处：《苏州大学》2016年博士论文　论文类型：学位论文

【摘要】：无监督学习是机器学习领域重要的研究方向之一,其应用非常的广泛。如数据聚类、复杂网络的主干网提取等。本文以投票集成聚类和复杂网络图聚类为切入点进行研究,取得的成绩包括:(1)针对数据的集成聚类问题,提出了基于扩展分类器系统的投票集成聚类方法。基于扩展分类器系统的投票集成聚类方法,首先利用扩展分类器系统在不同聚类个数的情况下生成一个聚类结果集合;然后引入分裂策略从所有候选值中确定聚类个数;最后,采用基于多数投票的一致性方法获得最终聚类结果。在人工数据和实际数据上的实验结果均表明了所提出方法的有效性。(2)在基于扩展分类器系统的投票集成方法的基础上,提出了基于扩展分类器系统的统一聚类集成框架。该框架包括了更多适用的融合准则、共识函数和自适应集成等内容。具体来说,在处理一个聚类任务的时候,所提出的方法首先会执行学习分类器系统来生成几个基聚类结果。为了使这些结果之间存在较大的多样性,本文对聚类数据使用不同的初始化,比如使用不同的聚类数目等。得到这些基聚类结果之后,我们提出的方法会使用相应的策略来生成最终的聚类结果。在人工数据和实际数据上的实验结果表明了所提框架的有效性。(3)针对复杂网络的图聚类问题,提出了一种基于不完全信息的无监督学习的复杂网络主干网提取方法。主干网提取的目的主要是压缩复杂网络的边和点数量,以尽量精简的子网络保留原网络的重要特征(如拓扑结构、点重要性特征等),从而帮助人们以更简单的形式来理解网络系统。本文以零模型为参考优化边过滤条件,并设计一种局部搜索模型。在四个真实网络上的实验结果表明本文所提出方法不仅大幅度减少了主干网中的离群点、而且更好地保留了原网络的各种特征、且比同类方法更加高效。
[Abstract]:Unsupervised learning is one of the important research directions in the field of machine learning. It is widely used, such as data clustering, backbone network extraction of complex networks, etc. The achievements of this paper include: (1) aiming at the problem of data clustering, an extended classifier system based voting ensemble clustering method is proposed, and an extended classifier system based voting ensemble clustering method is proposed. First, an extended classifier system is used to generate a set of clustering results under different clustering numbers; then a split strategy is introduced to determine the number of clusters from all candidate values. The final clustering results are obtained by using the consensus method based on majority voting. The experimental results on both artificial and actual data show that the proposed method is effective and based on the extended classifier system based voting integration method. A unified clustering integration framework based on extended classifier system is proposed. The framework includes more applicable fusion criteria, consensus functions and adaptive integration. The proposed method will first perform a learning classifier system to generate several base clustering results. In order to make these results more diverse, the clustering data are initialized differently in this paper. For example, using different clustering numbers, and so on. After obtaining these base clustering results, The proposed method will use the corresponding strategy to generate the final clustering results. The experimental results on artificial data and actual data show the effectiveness of the proposed framework. In this paper, an unsupervised learning method based on incomplete information is proposed to extract the backbone of complex networks. The main purpose of backbone network extraction is to compress the number of edges and points of complex networks. In order to help people understand the network system in a simpler form, the important features of the original network (such as topological structure, pointwise feature, etc.) are preserved by the subnetwork as concise as possible. In this paper, the zero model is used as the reference to optimize the edge filtering condition. Experimental results on four real networks show that the proposed method not only greatly reduces outliers in the backbone network, but also preserves all kinds of features of the original network. And more efficient than the same method.
【学位授予单位】：苏州大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP181;O157.5

【参考文献】