Fig. 2

a A simplified example of a 2-D model space for variables X1 and X2. The recursive splitting of the model space identifies three splits in a sequential manner that optimally partition the data to minimize the prediction error for an outcome, Y. b The corresponding decision tree displays the split points (in A) as internal nodes in the tree. At each node, a binary question is asked, which in the continuous case, results in subsetting the variable range. The terminal nodes for a regression tree represent non-overlapping regions in the model space. The label given to a region is simply the majority class of the outcome, Y, in the region. Predictions for a new sample are obtained by tracing the new sample down the decision tree into a terminal region with a given label