Improved SVM for Learning Multi-Class Domains with ROC Evaluation

The area under the ROC curve (AUC) has been used as a criterion to measure the performance of classification algorithms even the training data embraces unbalanced class distribution and cost-sensitiveness. Support vector machine (SVM) is accepted to be a good classification algorithm in classification learning. This paper describes an improved SVM learning method, where RBF is used as its kernel function, and the parameters of RBF are optimized by genetic algorithm. Within the parameter optimization and SVM learning, AUC is used as the evaluation criterion. The improved method can be used to deal with multi-class classification domains. Compared to the previous SVM algorithm, the improved SVM appears to have better learning performance.