Abstract

Often it is necessary to have an accurate estimate of the probability that a classifier prediction is indeed correct. Many classifiers output a prediction score that can be used as an estimate of that probability but for many classifiers these prediction scores are not well calibrated. If enough training data is available, it is possible to post process these scores by learning a mapping from the prediction scores to probabilities. One of the most used calibration algorithms is isotonic regression. This kind of calibration, however, requires a decent amount of training data to not overfit. But many real world data sets do not have excess amount of data that can be set aside for calibration. In this work, we have developed a data generation algorithm to produce more data from a limited sized training data set. We used two variations of this algorithm to generate the calibration data set for isotonic regression calibration and compared the results to the traditional approach of setting aside part of the training data for calibration. Our experimental results suggest that this can be a viable option for smaller data sets if good calibration is essential.