Probabilistic Random Forests: Predicting Data Point Specific Misclassification Probabilities

Recently proposed classification algorithms give estimates or worst-case bounds for the probability of misclassification [Lanckriet et al., 2002][L. Breiman, 2001]. These accuracy estimates are for all future predictions, even though some predictions are more likely to be correct than others. This paper introduces Probabilistic Random Forests (PRF), which is based on two existing algorithms, Minimax Probability Machine Classification and Random Forests, and gives data point dependent estimates of misclassification probabilities for binary classification. A PRF model outputs both a classification and a misclassification probability estimate for the data point. PRF makes it possible to assess the risk of misclassification, one prediction at a time, without detailed distribution assumptions or density estimation. Experiments show that PRFs give good estimates of the error probability for each classification.