TUNING SVM PARAMETERS FOR CLASSIFYING GEOGRAPHICAL ORIGINS OF CHINESE MEDICAL HERBS.

The quality of Chinese medical herbs depends on their geographical origins. Combining support vector machine (SVM) with Fourier transform infrared (FTIR) spectroscopy is a fast, accurate and nondestructive technique to classify the herbal origins. It is time consuming to tune the key parameters of SVM elaborately. However, a special leave-one-out (LOO) error bound of generalization ability (i.e., the ratio of the number of support vectors to the number of training vectors) can be estimated through training binary SVM once. In this paper, averaging all LOO error bounds over binary SVM classifiers is defined as the LOO error bound for multiclass problem according to one-against-other technique. Using this error bound and grid search, we can estimate the regularization and kernel parameters in SVM efficiently. In our experiments, 269 spectral curves of Dahurian Angelica Root (Baizhi in Chinese) from four provinces in China are examined in detail using polynomial and RBF kernels. On eight testing subsets, the average error rate is 3.53%. For four geographical origins, the highest error rate 6.07% is obtained using Zhejiang's herb. These results demonstrate that for classifying the herbal origins our proposed method can effectively tune SVM parameters and achieve high classification accuracy simultaneously. [ABSTRACT FROM AUTHOR]