Symbolically Aligning Observed and Modelled Behaviour

Abstract

Conformance checking is a branch of process mining that aims to assess to what degree a given set of log traces and a corresponding reference model conform to each other. The state-of-the-art approach in conformance checking is based on the concept of alignments. Alignments express the observed behaviour in terms of the reference model while minimizing the number of mismatches between the event data and the model. The currently known best algorithm for constructing alignments applies the A* shortest path algorithm for each trace of event data. In this work, we apply insights from the field of model checking to aid conformance checking. We investigate whether alignments can be computed efficiently via symbolic reachability with decision diagrams. We designed a symbolic algorithm for computing shortest-paths on graphs restricted to 0- and 1-cost edges (which is typical for alignments). We have implemented our approach in the LTSmin model checking toolset and compare its performance with the A* implementation supported by ProM. We generated more than 4000 experiments (Petri net model and log trace combinations) by setting various parameters, and analysed performance and related these to structural properties. Our empirical study shows that the symbolic technique is in general better suited for computing alignments on large models than the A* approach. Our approach is better performing in cases where the size of the state-space tends to blow up. Based on our experiments we conclude that the techniques are complementary, since there is a significant number of cases where A* outperforms the symbolic technique and vice versa.