Using unsupervised analysis to constrain generalization bounds for support vector classifiers