Mining discriminative class codes for multi-class classification based on minimizing generalization errors
Mongkon Eiadon, Luepol Pipanmaekaporn, Suwatchai Kamonsantiroj · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2016
Error Correcting Output Code (ECOC) has emerged as one of promising techniques for solving multi-class classification. In the ECOC framework, a multi-class problem is decomposed into several binary ones with a coding design scheme. Despite this, the suitable multi-class decomposition scheme is still ongoing research in machine learning. In this work, we propose a novel multi-class coding design method to mine the effective and compact class codes for multi-class classification. For a given n-class problem, this method decomposes the classes into subsets by embedding a structure of binary trees. We put forward a novel splitting criterion based on minimizing generalization errors across the classes. Then, a greedy search procedure is applied to explore the optimal tree structure for the problem domain. We run experiments on many multi-class UCI datasets. The experimental results show that our proposed method can achieve better classification performance than the common ECOC design methods.