Probability Estimation via Error - Correcting Output

Previous research has shown that a technique called error-correcting output coding (ECOC) can dramatically improve the classiication accuracy of supervised learning algorithms that learn to classify data points into one of k 2 classes. In this paper, we will extend the technique so that ECOC can also provide class probability information. ECOC is a method of converting k-class supervised learning problem into a large number L of two-class supervised learning problems and then combining the results of these L evaluations. The underlying two-class supervised learning algorithms are assumed to provide L probability estimates. The problem of computing class probabilities is formulated as an over-constrained system of L linear equations. Least squares methods are applied to solve these equations. Accuracy and reliability of the probability estimates are demonstrated.