Fast Opposite Maps: An Iterative SOM-Based Method for Building Reduced-Set SVMs

Abstract

Opposite Maps (OM) is a method that can be used to induce sparse SVM-based and LS-SVM-based classifiers. The main idea behind the OM method is to train two Self-Organizing Maps (SOM), one for each class, $\mathcal{C}_{-1}$ and $\mathcal{C}_{+1}$, in a binary classification context and then, for the patterns of one class, say $\mathcal{C}_{-1}$, to find the closest prototypes among those belonging to the SOM trained with patterns of the other class, say $\mathcal{C}_{+1}$. The subset of patterns mapped to the selected prototypes in both SOMs form the reduced set to be used for training SVM and LSSVM classifiers. In this paper, an iterative method based on the OM, called Fast Opposite Maps, is introduced with the aim of accelerating OM training time. Comprehensive computer simulations using synthetic and real-world datasets reveal that the proposed approach achieves similar results to the original OM, at a much faster pace.