Similarity measures are commonly applied for a variety of process mining techniques, such as trace clustering, conformance checking, and event abstraction. Yet, these measures generally fail to recognize similarity based on structural process features, such as the order of activities, loops, skips, choices, and parallelism. To make this more explicit, we propose a set of properties that allow to evaluate, what kind of structural features are reflected by a similarity measure. We further propose a novel approach leveraging existing graph-based algorithms and instance graphs to extract high-level structural features (loops, skips, choices, and parallelism) from traces, such that they can be used to extend and improve existing similarity measures. These algorithms are well-established in graph theory and can be computed efficiently. Finally, we provide an evaluation of the proposed approach based on synthetic and real-world datasets. The evaluation provides evidence that the additional graph-based features can substantially improve the similarity comparison of traces in several cases. This applies in particular for the comparison of user behavior (e.g., based on eye tracking data) where structural features enable the detection of specific behavioral patterns.
Schreiber et al. (Thu,) studied this question.