基于改进的SVM学习算法及其在信用评分中的应用  

An improved SVM learning algorithm and its applications to credit scorings

作　　者：陆爱国[1] 王珏[2] 刘红卫[1] 

机构地区：[1]西安电子科技大学理学院,西安710071 [2]中国科学院数学与系统科学研究院预测科学研究中心,北京100190

出　　处：《系统工程理论与实践》2012年第3期515-521,共7页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(70801058;70971052;61072144)

摘　　要：对于处理大规模问题的信用评分方法除要求达到一定的准确率之外,其速度、可解释性、简洁性等性能也非常重要.借鉴SMO的思想,首先提出一个基于三变量的改进的SVM学习算法,即将SVM问题分解为一系列含有三个变量的二次规划子问题,其优点是所求的相应松弛子问题都有解析解,使得该方法能够更加精确和快速地逼近最优解;其次将新算法应用于信用评分问题,在UCI机器学习库中的三个公共数据集上的数值试验表明了新方法的有效性:不仅节省了模型的计算代价,而且还提高了分类精度.A credit scoring method for a large problem not only achieves a certain its speed, interpretability, simplicity and other performance are also very important accuracy In this paper, a novel method called an improved SVM learning algorithm based on three-variable working set （ISVM-TV） is presented. This algorithm is derived by solving a series of the QP problems with only three points and the corresponding relaxation subproblems are solved analytically so that the proposed method approaches to the optimal solution more quickly. The proposed method is introduced to credit scoring and three datasets from UCI machine learning datasets are selected to demonstrate the method＇s competitive performance. Moreover, ISVM-TV shows a superior performance in saving the computational cost and improving classification accuracy.

关 键 词：支持向量机 三变量工作集 序列最小优化法 最大违背对 信用评分 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]