• There are many mature algorithms for creating decision trees. We used an off-the-shelf implementation of the C4.5 algorithm [15] implemented in the Weka [16] toolkit. The type of trees we chose required discrete rather than continuous coverage values. We experimented with several ways of dividing coverage into nominal values and found that a simple two-way division into high and low coverage yielded the best results. This was a natural choice since as Figure 3 shows, there are many more methods with 0% and 100% coverage than methods with coverage in the noninclusive range (0%, 100%). Furthermore, we found that the “split point” had little effect on the success rate, though higher splits were marginally better.