Imperative versus Declarative Process Mining: An Empirical Comparison

Process modelling notations fall in two broad categories: declarative notations, which specify the rules governing a process; and imperative notations, which specify the flows admitted by a process. We study the question whether certain process logs are better suited for mining to imperative than declarative notations. We attack this question by applying a flagship imperative and declarative miner to a standard collection of process logs, then evaluate the quality of the output models wrt. the standard output model metrics of precision and generalisation. This approach requires perfect fitness of the output model, which substantially narrows the field of available miners; we use the Inductive Miner and Minerful. With the metrics in hand, we statistically evaluate the hypotheses that (1) one miner consistently outperforms the other on one of the metrics, and (2) there exist subsets of logs more suitable for imperative respectively declarative mining. We confirm both hypotheses: (1) declarative output models have statistically significant higher generalisation (with no significant difference on precision), and (2) we find subsets of logs for which the imperative respectively declarative miner significantly outperforms the other.