基于ROC曲线寻优的支持向量机性能研究

支持向量机在小样本模式识别领域具有优势,但其性能评估及核参数、正则化参数的选择尚未有标准算法.将受试者操作特性曲线(Receiver Operating Characteristic,ROC)引入支持向量机分类性能分析和建模参数优化问题.在核参数及正则化参数所构成的二维空间中,调整模型参数阈值描绘ROC曲线,通过比较不同分类器ROC曲线下面积实现模型的性能分析,研究了基于ROC曲线最佳工作点的模型优化问题.工程实例表明,ROC曲线下面积有效地量化了模型的识别性能,并给出了一定寻优范围内的模型参数最优点,可以在SVM模型参数优化问题中推广应用.

Performance Evaluation with Optimization Strategy for Support Vector Machine Based on ROC Curve

Support vector machine (SVM) has become a popular tool in the area of pattern recognition, and parameters selection for SVM is an important issue to make it practically useful. In this paper, we introduced Receiver Operating Characteristic Curve into the performance evaluation and model optimization of SVM within the kernel parameters s and penalty factor c. Area under ROC curve was applied to the model evaluation, and model optimization was performed by seeking of optimal operating point of ROC. Pattern recognition experiment with UCI dataset shows that ROC curve is an effective approach for performance evaluation and optimization of SVM.