Can under-exploited structure of original-classes help ECOC-based multi-class classification?

Error correcting output codes (ECOC) is a popular framework for addressing multi-class classification problems by combing multiple binary sub-problems. In each binary sub-problem, at least one class is actually a “meta-class” consisting of multiple original classes, and treated as a single class in the learning process. This strategy brings a simple and common implementation of multi-class classification, but simultaneously, results in the under-exploitation of already-provided structure knowledge in individual original classes. In this paper, we present a new methodology to show that the utilization of such prior structure knowledge can further strengthen the performance of ECOCs, and the structure knowledge is formulated under the cluster and manifold assumptions, respectively. Finally, we validate our methodology on both toy and real benchmark datasets (UCI, face recognition and objective category), consequently validate the structure knowledge of individual original classes for ECOC-based multi-class classification.