Conformal ECOC Machines

Conformal prediction (CP) has received a signif-icant attention in the last decade due to its capability of pro-viding confidences for class predictions. However, the literature lacks of a computationally efficient implementation of CP for multi-class reduction techniques such as one-versus-all, ECOC etc. To address this issue we propose two implementations of CP for multi-class reduction: Conformal ECOC Machine (cECOC) and Conformal Poisson ECOC Machine (cpECOC). We show that both machines are computationally efficient and are capable of working with any reduction techniques based on binary code matrices. Moreover, we show that the second machine, cpECOC, probabilistically incorporates the error correction property of ECOC into CP framework using the Poisson binomial distribution. Conducted experiments on UCI datasets demonstrate that both machines output valid and efficient prediction sets.