Title: A note on ROC-optimizing support vector machines
Authors: Seung Jun Shin - Korea University (Korea, South) [presenting]
Abstract: Unbalanced classification where one class dominates another is frequently-encountered in practice. Most classifiers that target to reduce misclassified examples may fail in such case. The ROCSVM directly optimize the AUC of ROC can be used as a natural alternative in the unbalanced classification, since AUC is a performance measure independent of threshold value that controls balance of two classes. We present some results about ROC SVM. First, we establish the piecewise linearity of the ROC-SVM solution and develop an efficient algorithm to recover entire trajectories of the solutions. Second, we develop the SCAD-penalized ROC-SVM to select informative variables in unbalanced classification.