Interpolation Kernel Machines: Reducing Multiclass to Binary

Interpolating classifiers interpolate all the training data and thus have zero training error. Recent research shows their fundamental importance for high-performance ensemble techniques and other advantages. Interpolation kernel machines belong to the class of interpolating classifiers and do generalize well. They have been demonstrated to be a good alternative to support vector machines. In this work we further improve their performance. We propose not to use their inherent multiclass classification capacity, but instead apply them for solving binary classification instances based on a mutliclass-to-binary reduction. We experimentally study this ensemble approach in combination with six reducing multiple-to-binary methods. The experimental results show that the one-versus-one scheme consistently demonstrates superior performance.