Abstract-and-compare stochastic conformance checking

Conformance checking is a key area of process mining focused on identifying and quantifying deviations between observed and modeled behavior. Traditional techniques only consider the frequency information in event logs, leading to one-sided analyses. In recent years, stochastic conformance checking, which considers the stochastic perspective of process models, has gained attention. However, state-of-the-art stochastic conformance checking techniques are often either not robust to partial mismatches or cannot be efficiently computed for broad classes of process models or large input sets, limiting their practical applicability. To address these challenges, we propose an abstract-and-compare approach, where stochastic languages are first abstracted and then compared using existing stochastic conformance checking techniques. Specifically, we introduce the stochastic Markovian abstraction, based on expected subtrace frequencies, and demonstrate how to compute it for bounded, livelock-free stochastic labeled Petri nets- a widely used stochastic process model formalism. The Markovian abstraction defines a stochastic language and can be combined with state-of-the-art stochastic conformance measures to obtain “new” derived measures. We evaluate two such derived measures on synthetic and real-world datasets, showing that the abstraction can effectively approximate model behavior, is computationally efficient, and improves robustness to partial mismatches. Additionally, we demonstrate how to obtain stochastic conformance diagnostics for the derived measures.