使用DBSCAN的FCM神经网络分类器  

FCM Neural Network Classifier Using Density-Based Spatial Clustering of Applications with Noise

作　　者：张晓倩[1] 杨波[1] 王琳[1] 梁志锋[1] 

机构地区：[1]济南大学山东省网络环境智能计算技术重点实验室,济南250022

出　　处：《模式识别与人工智能》2016年第2期185-192,共8页Pattern Recognition and Artificial Intelligence

基　　金：国家科技支撑计划项目(No.2012BAF12B07-3);国家自然科学基金项目(No.81301298;61302128;61373054;61203105;61173078;61173079;61070130);山东省自然科学基金项目(No.ZR2012FQ016;ZR2012FM010;ZR2011FZ001;ZR2011FL022;ZR2010FM047;ZR2010FQ028);济南市青年科技明星计划项目(No.2013012)资助~~

摘　　要：针对浮动质心法(FCM)在实现过程采用的K-means算法不易发现任意形状簇及对离群点敏感等缺陷,提出使用具有噪声的基于密度的聚类算法(DBSCAN)改进FCM神经网络分类器的方法.DBSCAN将离群点看作无法处理的点,并能发现任意形状的簇,将分区空间中的染色点划分成若干个更准确的分区.此外,定义优化目标函数,并用粒子群优化算法优化神经网络的各个参数,获得最优的分类模型.在UCI数据库上的对比实验表明,改进后的FCM方法在分类精度、鲁棒性和运行时间方面均优于原有FCM.Arbitrary shape clusters are difficult to be found by the K-means algorithm adopted in the implementation of floating centroids method （FCM）. Moreover, K-means algorithm is sensitive to outliers. Aiming at these problems, density-based spatial clustering of applications with noise （DBSCAN） algorithm is used to improve FCM neural network classifier in this paper. The outliers are considered as the points that can not be dealt with, and clusters of arbitrary shape can be found by DBSCAN algorithm. Thus, the color points in the partition space can be divided into several more accurate partitions. In addition, an optimization objective function is defined, and the particle swarm optimization algorithm is employed to optimize the parameters of the neural network to obtain an optimal classification model. Several commonly used datasets from UCI database are selected to conduct a comparative experiment. The experimental results show that the improved FCM method generates better performance on classification accuracy, robustness and running time than that of the original FCM method.

关 键 词：神经网络 浮动质心法(FCM) 分区空间 具有噪声的基于密度的聚类算法(DBSCAN) 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论] TP183[自动化与计算机技术—计算机科学与技术]