Python Process Mining — Core Concepts

What Process Mining Is

Process mining is a family of techniques that extract knowledge about business processes from event logs. Unlike traditional business process modeling (where humans draw diagrams of how work should flow), process mining discovers how work actually flows based on recorded data.

The field was pioneered by Wil van der Aalst at Eindhoven University and has grown into a major industry, with companies like Celonis, UiPath, and Minit offering commercial tools. Python’s pm4py library brings these capabilities to developers.

The Three Types of Process Mining

1. Process Discovery

Input: event log. Output: a process model (Petri net, BPMN diagram, or directly-follows graph).

The algorithm reads sequences of activities and constructs a model that explains the observed behavior. No pre-existing model is needed.

2. Conformance Checking

Input: event log + process model. Output: deviations.

Compares what happened (the log) against what should happen (the model). Identifies cases that skip steps, repeat steps, or take unexpected paths.

3. Enhancement

Input: event log + process model. Output: improved model.

Enriches the model with performance data (bottleneck detection, waiting times) or extends it with additional perspectives (resources, costs).

Event Logs: The Raw Material

An event log is a collection of cases, each containing a sequence of events. Every event has at minimum:

Field	Example	Purpose
Case ID	”ORD-4521”	Groups events belonging to the same process instance
Activity	”Approve Request”	What happened
Timestamp	”2026-03-15 09:23:00”	When it happened

Optional but valuable fields:

• Resource — who performed the activity
• Cost — how much the activity cost
• Custom attributes — department, priority, region

Example log fragment:

Case ID	Activity	Timestamp	Resource
C001	Submit Request	2026-03-01 09:00	Alice
C001	Review Request	2026-03-01 11:00	Bob
C001	Approve Request	2026-03-01 14:00	Carol
C002	Submit Request	2026-03-01 09:30	Dave
C002	Review Request	2026-03-01 10:00	Bob
C002	Reject Request	2026-03-01 10:30	Carol

Discovery Algorithms

Alpha Miner

The original algorithm. Looks at ordering relationships between activities (A always before B, C and D happen in parallel) to construct a Petri net. Simple but struggles with noise and complex patterns.

Inductive Miner

Recursively splits the log into sub-logs and constructs a process tree. Much more robust than Alpha — handles noise, incomplete logs, and complex patterns. The recommended starting point.

Heuristics Miner

Uses frequency-based thresholds to handle noise. If activity A is followed by B in 95% of cases and by C in 5%, it can filter out the rare path as noise. Good for messy real-world logs.

Directly-Follows Graph (DFG)

Not a formal discovery algorithm, but the simplest visualization. Shows which activities directly follow which, with frequency counts on edges. Fast to compute but can be misleading for concurrent processes.

Conformance Checking Methods

Token-based replay: Replays each case on the Petri net model, counting “missing tokens” (the model couldn’t produce the observed activity) and “remaining tokens” (the case ended before the model expected).

Alignment-based: Finds the optimal alignment between the observed trace and the model. More accurate but computationally expensive. Identifies exactly where each case deviates.

Metrics:

• Fitness — what fraction of log behavior is captured by the model? (1.0 = perfect fit)
• Precision — does the model allow behavior not in the log? (1.0 = no extra behavior)
• Generalization — will the model fit future unseen cases?
• Simplicity — is the model understandable?

These four compete with each other. A model that accepts everything has perfect fitness but zero precision. Balancing them is the art of process mining.

Python Library: pm4py

pm4py is the standard open-source process mining library for Python. Key capabilities:

• Read event logs (CSV, XES, Parquet)
• Discover process models (Alpha, Inductive, Heuristics miners)
• Conformance checking (token replay, alignments)
• Performance analysis (bottleneck detection, waiting times)
• Social network analysis (who works with whom)
• Visualization (Petri nets, BPMN, DFGs)

Common Misconception

“Process mining only works for simple, linear processes.” It actually excels at revealing complexity. The most valuable insights come from processes that seem simple but are actually chaotic — insurance claims, patient pathways, IT incident management. Process mining shows the spaghetti-like reality that no manual audit would uncover.

When to Use Process Mining

• After a system migration — verify the new system follows the same process
• Compliance audits — prove that required steps aren’t being skipped
• Performance optimization — find bottlenecks and rework loops
• RPA (Robotic Process Automation) preparation — understand the process before automating it
• Continuous monitoring — detect process drift over time

One thing to remember: Process mining reverses the traditional approach — instead of designing a process and hoping people follow it, you extract the real process from data and see what actually happens, gaps and all.