PROJECT TITLE :

In-Sample and Out-of-Sample Model Selection and Error Estimation for Support Vector Machines

ABSTRACT:

In-sample approaches to model choice and error estimation of support vector machines (SVMs) aren't as widespread as out-of-sample ways, where part of the data is faraway from the training set for validation and testing purposes, mainly as a result of their practical application is not straightforward and therefore the latter provide, in many cases, satisfactory results. During this paper, we tend to survey some recent and not-so-recent results of the info-dependent structural risk minimization framework and propose a proper reformulation of the SVM learning algorithm, thus that the in-sample approach will be effectively applied. The experiments, performed both on simulated and real-world datasets, show that our in-sample approach will be favorably compared to out-of-sample ways, particularly in cases where the latter ones give questionable results. In particular, when the amount of samples is tiny compared to their dimensionality, like in classification of microarray information, our proposal can outperform standard out-of-sample approaches like the cross validation, the leave-one-out, or the Bootstrap strategies.