Improving Text Classifier Performance based on AUC

To evaluate the performance of text classifiers, we usually look at measures related to precision and recall, and most machine learning methods are optimized for these measures. In recent year, the use of receiver operating characteristics (ROC) graph and its extension area under the ROC curve (AUC) in gauging classifier performance has attracted much attention from the machine learning community. This measure is especially useful when a data set is imbalanced or when operating characteristics are unknown. Some researchers have started investigating the optimization of existing learning model for this new performance criterion. In this paper, we proposed modifications to the well-known weight updating text classifier sleeping-experts (SE) for AUC optimization. Our experiments show that through our new sampling and updating strategy we can improve the classifier both in terms of AUC and the traditional performance measures