We study the problem of designing support vector machine (SVM)classifiers that minimize the maximum of the false alarm and missrates. This is a natural classification setting in the absence of priorinformation regarding the relative costs of the two types of errorsor true frequency of the two classes in nature. Examining two approaches– one based on shifting the offset of a conventionally trainedSVM, the other based on the introduction of class-specific weights –we find that when proper care is taken in selecting the weights, thelatter approach significantly outperforms the strategy of shifting theoffset. We also find that the magnitude of this improvement dependschiefly on the accuracy of the error estimation step of the trainingprocedure. Furthermore, comparison with the minimax probabilitymachine (MPM) illustrates that our SVM approach can outperformthe MPM even when the MPM parameters are set by an oracle.
