On Tuning Hyper-Parameters of Multiclass Margin Classifiers

The choice of hyper-parameters (e.g. kernel parameters) can significantly affect generalization performance of large margin classifiers. In this paper we are con- cerned with the problem of tuning these values in the case of multi-class problems that have been recast into a set of binary problems. We report several experimental results comparing independent and joint tuning of the hyper-parameters of the binary clas- sifiers. Several different encoding strategies are explored, including error correcting output codes. Tuning was carried out by using a validation set and a newly introduced bound on the leave-one-out error.