Conformance Checking

The second of the three pillars of process mining: given an event log and a process model, quantify and localise deviations (sources/2011-vanderaalst-process-mining-book, Ch. 8). Distinguished from discovery (log → model) and enhancement (log + model → improved model).

Three techniques

Token replay (§8.2) — push each trace through the model; count missing/remaining tokens. Simple, but gives coarse diagnostics on non-fitting traces.
Alignments (§8.3) — compute a cost-optimal synchronous product between trace and model (log moves / model moves / synchronous moves). Dominant technique since the late 2000s; produces interpretable step-by-step deviation diagnostics. Introduced to the textbook in the 2nd edition.
Footprint comparison (§8.4) — compare direct-succession matrices of log vs model; cell-level disagreements reveal structural mismatches. Fast but log-model-symmetric (no trace-level diagnostics).

Four quality dimensions

Conformance is scored along four dimensions (sources/2011-vanderaalst-process-mining-book, §6.4.3; echoed in sources/2018-dumas-fundamentals-of-bpm §11.4.4):

Fitness — can the model replay the traces in the log?
Precision — does the model only allow behaviour seen in the log, or does it over-generalise?
Generalisation — will the model handle unseen-but-plausible future behaviour?
Simplicity — Occam: is the model the smallest adequate explanation?

These trade off against each other; the four-way tension is the conformance equivalent of the Devil's Quadrangle in redesign.

Applications (§8.5)

Compliance auditing — check four-eyes principles, segregation of duties, SOX/GDPR rules.
Model repair — patch the model where it disagrees with observed reality.
Discovery algorithm evaluation — use alignment-based fitness/precision as benchmark metrics.
Model-to-log linkage — annotate log events with their model position for downstream analysis.

Two viewpoints on deviations

Per sources/2011-vanderaalst-process-mining-book §2.3:

Descriptive model (the model is wrong): deviations reveal model error; fix the model.
Normative model (reality is wrong): deviations reveal non-compliance; fix the process or the people.

The same alignment produces different actions depending on which viewpoint is adopted.

methods/process-mining-basics · concepts/process-discovery · concepts/process-model-quality · concepts/operational-support · frameworks/declare (declarative conformance via LTL replay) · entities/wil-van-der-aalst

Out-of-scope

Not to be confused with Statistical Process Control (SPC — control charts over numerical measurements, see sources/2007-slack-operations-management Ch. 17) or chemical-engineering process control (sources/2007-bao-lee-process-control-passivity). These are three distinct traditions that share the word "control".