Algorithm 1 presents the pseudo-code of COAL algorithm. As the first step, the
M5P algorithm is used to construct two diverse regression trees (say, m1 and m2) via
employing distinct parameters M1 and M2 respectively, where M1 and M2 determine the
minimal number of examples in each leaf of m1 and m2, respectively [Wang and Witten
1997]. Let L1 and L2 be the current labeled data sets of m1 and m2 respectively, and
L1 and L2 are initialized by the same labeled data set L. In each iteration, COAL uses
the PLE determined by m1 to augment the labeled set L2, and vice versa. After that,
we should replenish the pool U with unlabeled instances to size p for next iteration.
After using the latest labeled sets to update two regression trees, COAL selects the
unlabeled configuration with the largest disagreement to simulate, and updates the
regression trees using the newly labeled configuration. It is notable that the unlabeled
data selected for simulation is chosen from the entire unlabeled data set U instead of