Error-correcting output codes based ensemble feature extraction

This paper proposes a novel feature extraction method based on ensemble learning. Using the error-correcting output codes (ECOC) to design binary classifiers (dichotomizers) for separating subsets of classes, the outputs of the dichotomizers are linear or nonlinear features that provide powerful separability in a new space. In this space, the vector quantization based meta classifier can be viewed as an ECOC decoder, where each learned prototype of a class can be seen as a codeword of the class in the new representation space. We conducted extensive experiments on 16 multi-class data sets from the UCI machine learning repository. The results demonstrate the superiority of the proposed method over both existing ECOC approaches and classic feature extraction approaches. In particular, the decoding strategy using a meta classifier is shown to be more computationally efficient than the linear loss-weighted decoding in state-of-the-art ECOC methods.

► We propose a novel feature extraction method based on ensemble learning by ECOC. ► The extracted features provide high separability of classes. ► The vector quantization based meta learner can be viewed as ECOC recoding. ► The proposed method demonstrated superior classification performance.