Process model

What is a process model?

A process model is a simplified picture of how a process runs: which activities happen, in what order, where the choices and the parallel paths are. It can show how a process is meant to work, what behaviour is formally allowed, or what patterns were actually found in the data.

The model is not the process itself. People and systems do the work; the model helps you design it, discuss it, run it, or analyse it. A useful model matches its purpose, whether that is a workshop drawing or an executable definition.

Kinds of process model

A hand-drawn model captures the agreed way of working. BPMN is the common notation for this, with tasks, events, and gateways that business readers can follow.

An executable model carries enough precise logic for a workflow engine to run it. Error paths, correlation, and technical detail matter far more here than in a workshop sketch.

A discovered model is derived from an event log by process discovery. It reflects recorded behaviour, shaped by the data, filters, and algorithm settings you chose.

A Petri net is a formal notation with places, transitions, and tokens. It is good for analysing choice, synchronisation, and parallel behaviour precisely, which is why process mining research leans on it.

An order example

An order model starts at "order received". Credit check and stock check can then run in parallel, and only when both pass does "release for shipping" follow. If stock is short, the process waits or the order is cancelled. After shipping come "create invoice" and "receive payment", and a return can be a separate loop.

A flat diagram that lines the activities up one after another hides the parallel checks. A model that draws every technical API call becomes unreadable for business users. The right level of detail depends on the goal: discussion, automation, compliance, or performance analysis.

Discovering a model from an event log

Process discovery uses the recorded traces to derive a model that summarises what happened. Algorithms differ in how they handle repetition, rare paths, parallelism, and incomplete logs. A model that follows every rare trace exactly turns into unreadable spaghetti; a heavily simplified one can hide important exceptions.

Filters and thresholds change the result, so keep the log version, the algorithm, and the parameters together with the model. And have people who know the process check that the relationships are plausible, because a pattern can come from wrong timestamps or bad case correlation rather than real behaviour.

Model versus event log versus process map

An event log holds the concrete events per process instance. A process model abstracts away from those into possible or expected behaviour. A process map, often a directly-follows graph, shows which activities directly followed which and how often; it is intuitive but does not always make choices and parallelism formally clear. A formal model can allow behaviour that never appeared literally in the log, such as another valid interleaving of parallel tasks. That ability to generalise is useful, but it needs limits.

Fitness, precision, and readability

Fitness measures how well the traces in the log can be replayed by the model. High fitness means the recorded behaviour rarely conflicts with the model. Precision measures whether the model avoids allowing lots of extra paths that were never seen; a model where anything can happen has high fitness but explains little.

Generalisation and simplicity matter too. A workable model copes with new but valid behaviour and stays readable. No single metric picks the best model on its own, which is why discovery is a judgement call, not a button.

What to watch out for with process models

Record the owner, purpose, scope, version, and source, and keep as-is, to-be, and executable models apart. Change a normative model when policy changes, and check which conformance results shift as a result. Re-discover a data-driven model when the real behaviour changes for good. A tidy diagram without reliable event data or clear meaning behind it is still just a hypothesis.