Support vector machine classifiers built using imperfect training data

Abstract

This paper extends the utility of asymmetric soft margin support vector machines by analytically modeling imperfect class labeling in the training data. It uses Receiver Operating Characteristics computations to first establish the strong relationship between the support vector machines performance and its ability to classify examples correctly, even in the presence of misclassified training examples. It uses statistically designed experiments to reveal that misclassification also affects training quality, and hence performance, though not as strongly. Still, our results give strong support for ones striving to develop the best trained support vector machine that is intended to be utilized,
for instance, for medical diagnostics, as misclassifications shrink decision boundary distance and increase
generalization error. Also, this study asserts that real life costs of making wrong classification should be incorporated in the support vector machine design optimization objective.