Measuring how well a process model matches reality using replay techniques and alignment-based analysis.
Once a process model has been discovered or designed, a fundamental question arises: does the model accurately represent the real process? Conformance checking addresses this by comparing an event log against a process model and quantifying the degree of alignment between the two. This is critical for auditing, regulatory compliance, and continuous improvement — if a model does not reflect reality, decisions based on it may be flawed (Rozinat & Van der Aalst, 2008).
Conformance checking has direct practical applications. Internal audit teams use it to verify that operational processes follow documented controls. Compliance officers detect deviations from regulated procedures in healthcare or finance. Process improvement teams identify where the "official" process differs from what employees actually do — revealing whether the process design is impractical or whether training is needed.
Van der Aalst (2016) defines four dimensions for evaluating the quality of a process model against event log data:
These four dimensions involve inherent trade-offs. Increasing precision often reduces generalisation; maximising fitness may produce overly complex models. Effective conformance analysis balances these dimensions based on the use case — a compliance audit may prioritise fitness, while a discovery exercise may prioritise precision and simplicity.
Token-based replay is the classic technique for measuring fitness. Each trace in the event log is replayed on the Petri net model: tokens are produced and consumed as activities fire. When the model cannot consume a token (the activity is not enabled), a "missing token" is recorded. When the model has tokens remaining after replay completes, "remaining tokens" are counted. Fitness is computed from the ratio of consumed, produced, missing, and remaining tokens across all traces (Rozinat & Van der Aalst, 2008).
While conceptually elegant and computationally efficient, token-based replay has limitations. It produces only a single fitness score per trace and cannot pinpoint exactly where deviations occur. It also assumes the model is a Petri net, which may not suit all modelling notations.
Alignments address these limitations by computing an optimal mapping between each trace in the log and a path through the model. The cost of an alignment is the number of "moves" that deviate from the model — for example, executing an activity that the model does not allow (a log move) or skipping an activity that the model requires (a model move). The minimal total cost across all possible alignments gives a precise, step-level measure of conformance (Carmona et al., 2018).
Alignments provide richer diagnostic information than token replay: they reveal exactly which activities deviate and at what point in the process. This makes them the preferred method for detailed conformance analysis, though at a higher computational cost. Modern process mining tools such as Celonis and PM4Py implement alignment-based conformance checking for real-world event logs.
You are auditing a purchase-to-pay process where the documented policy requires: "Requisition → Approval → Purchase Order → Goods Receipt → Invoice → Payment." The event log shows that in 40% of cases, no Approval event occurs before the Purchase Order.
What does this tell you about the fitness and precision of the model? What would you recommend — enforcing the approval step, or updating the process model to reflect that approvals sometimes happen after the PO? How might an alignment help you diagnose individual cases?