Improving Classification Accuracy of Large Test Sets Using the Ordered Classification Algorithm

We present a new algorithm called Ordered Classification, that is useful for classification problems where only few labeled examples are available but a large test set needs to be classified. In many real-world classification problems, it is expensive and some times unfeasible to acquire a large training set, thus, traditional supervised learning algorithms often perform poorly. In our algorithm, classification is performed by a discriminant approach similar to that of Query By Committee within the active learning setting. The method was applied to the real-world astronomical task of automated prediction of stellar atmospheric parameters, as well as to some benchmark learning problems showing a considerable improvement in classification accuracy over conventional algorithms.