Using Unsupervised Analysis to Constrain Generalization Bounds for Support Vector Classifiers