项目名称: 基于三元粗糙输出编码的带自适应惩罚因子的支持向量机多分类模型研究
项目编号: No.61262047
项目类型: 地区科学基金项目
立项/批准年度: 2013
项目学科: 自动化技术、计算机技术
项目作者: 饶泓
作者单位: 南昌大学
项目金额: 35万元
中文摘要: 由于支持向量机(SVM)在处理高维小样本数据时的识别精度显著优于传统机器学习方法,因此基于SVM的多分类方法与应用是目前研究的热点。但是由于采用SVM进行多分类时必须将SVM由二分类扩展至多分类,易出现决策盲区、数据集倾斜,一致多样性等问题,从而导致最终决策产生偏差甚至是错误的情况。因此,为解决上述问题,本课题拟舍弃现有的编码思想,采用三元模型{+1,0,-1}取代传统的二元模型{+1,-1},通过快速获取样本简单超球信息,构建自适应的期望标签矩阵编码(拟命名为三元粗糙输出编码,Ternary Rough Output Codes, TROC)。其次,为减少高维特征空间分类边缘内不可分情况的出现,本课题引入信息熵和模糊数学理论研究自适应惩罚因子计算方法和不可分类的模糊判定方法,在这些研究的基础上,提出一种有效的基于三元粗糙输出编码的带自适应惩罚因子的SVM多分类模型,并研究其实际应用。
中文关键词: 多分类;三元编码矩阵;层次分析方法;支持向量机;数据补全策略
英文摘要: Because the Support Vector Machine (SVM)'s recognition accuracy is significantly better than the traditional machine learning methods' in processing small sample of high-dimensional data , SVM-based multi- class classification method and its application has been being current research hot spot in pattern recognition area. However,SVM-based multi-class classfication must extend the SVM two classification to multi-class classification, which is easily lead to the decision-making blind data ,set tilting ,consistent-diverse and other problems, and result in final decision-making bias, or even wrong decision. Therefore, in order to solve these problems, this project intends to abandon the existing coding idea, using three models {+1,0, -1} rather than the traditional binary model {+1, -1}, construct self-adaptive expectations label matrix code(called Ternary Rough Output Codes,TROC) by quickly obtaining the sample's hyperspherical information. And then, to reduce inseparable situation appears in high dimensional feature space classification edge,this project will adopt the theory of information entropy and fuzzy mathematical to research self-adaptive penalty factor calculation methods and fuzzy judgment method for inseparable labels. Finally, on the basis of above studies,this project will propose a more
英文关键词: mulit-class classification;ternary code matrix;hierarchy analysis method;SVM;Strategy of data completion.