A Dynamic Ensemble Selection Strategy for Improving Error Correcting Output Codes Algorithm

Abstract

Error correcting output codes (ECOC) is an effective approach for the multiclass classification problem by decomposing a multiclass problem to a set of binary class problems. Up to now, most of proposed ECOC algorithms mainly focus on the generation of codematrices, while this study proposes a Dynamic Ensemble Selection strategy for improving the performance of the original ECOC algorithms. In this strategy, a codematrix is generated by an ECOC algorithm firstly, and then some feature selection algorithms are deployed to select feature subsets for each column in the codematrix. When an unknown sample arrives, the Fisher criterion is deployed to select the best feature subset from the candidate subsets to match a column. After determining the optimal feature subsets for various columns, the distance evaluation process is carried out to assign the sample to a class. The proposed strategy is tested on UCI data sets with three classical ECOC algorithms. Experiment results prove that our strategy can improve the performance of the original ECOC algorithms in most cases.