An Improved SMO Algorithm for Credit Risk Evaluation
主讲人：王珏（预测科学研究中心） 时间：2015年10月15日上午9:00 地点：N210
【Abstract】Sequential minimal optimization (SMO) is the most commonly used algorithm for numerical solution of SVM, but traditional SMO is quite limited to the long execution time because of its high computational complexity. We present an improved SMO learning algorithm named FV-SMO in this paper. At each iteration, it jointly optimizes four variables and an theorem is proposed to guarantee an analytical solution of subproblem. Three credit datasets are selected to demonstrate the performance of FV-SMO, including China credit dataset and two public datasets: German credit dataset from UCI and Darden credit dataset from CD-ROM databases. China credit dataset is generated based on a multi-dimensional and multi-level credit risk indicator system of China credit data. Experimental results demonstrate that FV-SMO is competitive in saving the computational cost and performs best in credit risk evaluation accuracy compared with other five popular classification methods.