Decision tree training is a highly unstable procedure. Small differences in the training set can cause vastly different decision tree structures, and often very different outcomes. Also, because they are piecewise-constant approximations, observations on the boundaries between regions are more likely to have high error rates.