Improving Generalization Abilities of Maximal Average Margin Classifiers

Maximal average margin classifiers (MAMCs) maximize the average margin without constraints. Although training is fast, the generalization abilities are usually inferior to support vector machines (SVMs). To improve the generalization abilities of MAMCs, in this paper, we propose optimizing slopes and bias terms of separating hyperplanes after the coefficient vectors of the hyperplanes are obtained. The bias term is optimized so that the number of misclassifications is minimized. To optimized the slope, we introduce a weight to the average of mapped training data for one class and optimize the weight by cross-validation. To improve the generalization ability further, we propose equally constrained MAMCs and show that they reduce to least squares SVMs. Using two-class problems, we show that the generalization ability of the unconstrained MAMCs are inferior to those of the constrained MAMCs and SVMs.