An Approximate Inductive Miner

Process discovery algorithms extract process models from business process event logs. Existing discovery algorithms require upfront filtering, or specific parameter input, to produce models with balanced quality dimensions on real-life event logs. We propose the Approximate Inductive Miner (AIM) to fill this gap and offer an automated way to discover sound models in polynomial time complexity, without any pre-processing or mandatory parameter input. AIM uses the existing Inductive Miner framework and applies clustering techniques to recursively identify structures in the event log. It additionally performs an approximate parameter optimisation to dynamically suggest a suitable parameter. We compare AIM with existing discovery algorithms on synthetic and real-life event logs, and evaluate the quality of the integrated parameter suggestion. We find that AIM on its own produces sound models with low control flow complexity and high precision, even on complex event logs. Additionally, AIM is able to handle a vast range of event log properties, such as infrequent and incomplete behaviour, without requiring any human parameter input or upfront filtering.