A genetic inspired optimization for ECOC

Abstract

In this work, we propose a novel Genetic Inspired Error Correcting Output Codes (ECOC) Optimization, which looks for an efficient problem-dependent encoding of the multi-class task with high generalization performance. This optimization procedure is based on novel ECOC-Compliant crossover, mutation, and extension operators, which guide the optimization process to promising regions of the search space. The results on several public datasets show significant performance improvements as compared to state-of-the-art ECOC strategies. © 2012 Springer-Verlag Berlin Heidelberg.