Research Of Classification Algorithm Evaluation Based On ROC

Classification is the most prominent branch of machine learning. The predictive ability of the algorithm is typically measured by its accuracy on the testing examples in the past. However,In the case of class skew or cost sensitive,accuracy is not enough. ROC(Receiver Operating Characteristic) has been used as a standard tool for the analysis and comparison of classifiers when the dataset has unbalanced classes or the cost of misclassification is unknown. Researchers have been examing ROC for two class classifications,but doing little work for that of more than two class ones.We proposd an effective way to approach multiclass ROC analysis method named EMAUC. The method applies binary algorithms to deal with multiclass learning problem by employing error correcting output coding. EMAUC has competive performance, better comprehension and irrelated with skew dataset. EMAUC is developed based on machine learning platform WEKA and ROCon.The system is evaluated with datasets from UCI data repository. Empirical results demonstrate that EMAUC is effective to evaluate classifiers.