Comprehensible Predictive Models for Business Processes

Table A1 summarizes the various characteristics of the synthetic models used in the experiments, including the number of event types, the size of the state space, whether a challenging construct is contained (loops, duplicates, nonlocal choice, and concurrency), and the entropy of the process defined by the model (estimated based on a sample of size 10,000). The original models may contain either duplicate tasks (two conceptually different transitions with the same label) or invisible tasks (transitions that have no label, as their firing is not recorded in the event log). We transformed all invisible transitions to duplicates such that, when there was an invisible task i in the original model, we added duplicates for all transitions t that, when fired, enable the invisible transition. These duplicates emulate the combined firing of t and i. Since we do not distinguish between duplicates and invisible tasks, we combined this category.