Evaluating the generalization ability of support vector machines through the bootstrap

The well-known bounds on the generalization ability of learning machines, based on the Vapnik-Chernovenkis (VC) dimension, are very loose when applied to Support Vector Machines (SVMs). In this work we evaluate the validity of the assumption that these bounds are, nevertheless, good indicators of the generalization ability of SVMs. We show that this assumption is, in general, true and assess its correctness, in a statistical sense, on several pattern recognition benchmarks through the use of the bootstrap technique.