Cost-Sensitive Ensemble of Support Vector Machines for Effective Detection of Microcalcification in Breast Cancer Diagnosis

This paper presents a new approach for the cost-sensitive classification problems based on the Boosting ensemble of support vector machines (SVMs). Different from conventional Boosting ensemble learning methods that adjust the distribution of training instances for minimizing the misclassification rate, the presented approach adjusts the training data distribution so as to minimize the expected cost of classification. This approach has been applied successfully in Microcalcification (MC) detection which is a typical cost-sensitive classification problem in breast cancer diagnosis. Its performance is evaluated by means of Receiver Operating Characteristics (ROC) curves and the expected costs of classification. Experimental results have consistently confirmed that the ROC of the SVM ensemble classifier is very close to the curve enveloping the base classifier ROC curves. This characteristic illustrates that the SVM ensemble is able to always improve the performance of the classification. Furthermore, the experimental results demonstrate that the method presented is able to not only increase the area under the ROC curve (AUC) but also minimize the expected classification cost.