Solving multiclass classification problems using Error Correcting Output Code (ECOC) schemes offers flexibility and robustness, but traditional approaches based on Rademacher sequences suffer from an exponential increase in dichotomies with the number of classes. This paper introduces two main contributions. First, we propose the use of Walsh sequences for directly constructing ECOC dichotomies, enabling multiclass problems with C classes to be addressed using only C − 1 classifiers, while preserving high pairwise Hamming distances between class codes ( C 2 ). Second, we present a principled method to estimate posterior class probabilities from the outputs of discriminative classifiers trained with Bregman divergences. This provides an alternative to Hamming decoding, improves confidence estimation, and avoids matrix inversion due to the orthogonality of Walsh codes. Experimental results on five multiclass datasets—including Fashion-MNIST, Letter Recognition, and Vowel—demonstrate consistent improvements in classification accuracy (e.g., a 5% error reduction on the Letter dataset) and computational efficiency over baseline methods such as OvR and ECOC-Rademacher. Performance was evaluated using standard classification accuracy and confusion matrices, validating the advantages of our approach in both predictive performance and scalability. • Employing Walsh sequences to create ECOC dichotomies. • Applicable to multiclass classification with a large number of classes. • Only C - 1 binary classifiers are required for C classes using Walsh sequences. • A principled formulation to estimate the posterior class probabilities. • Extensive experiments support the effectiveness of the proposal.
Álvarez-Pérez et al. (Tue,) studied this question.