Figure 4.

Running time (in seconds) of each algorithm and each model selection strategy on the benchmark datasets. For most datasets, using the sample compression bounds as a model selection strategy considerably reduces the running time compared to ten-fold cross-validation. SCM tends to run faster than CART, which was expected given that for models of equal depth, CART must evaluate many more rules than SCM. Of note, the time is shown on a logarithmic scale and thus, a difference in one unit corresponds to a running time that is ten times smaller.