PM: Methods in pm4py

Here’s a comparison table of the main methods in process mining (as available in PM4Py) so you can see their differences at a glance:

Process Discovery Methods

Method	Output Model	Pros	Cons	Best Use Case
Alpha Miner	Petri Net	Simple, foundational, easy to explain	Very sensitive to noise/incomplete logs	Educational/demo purposes, very clean logs
Heuristics Miner	Heuristics Net / Petri Net	Handles noise, considers frequency	May oversimplify rare behavior	Real-life logs with noise and high variability
Inductive Miner	Petri Net / Process Tree / BPMN	Always produces sound models, block-structured	May abstract away some detail	General-purpose discovery, recommended default
ILP Miner	Petri Net	Precise, mathematically grounded	Heavy computational cost	Small/medium logs where precision is critical
DFG Discovery	Directly-Follows Graph	Very fast, intuitive visualization	Lacks formal semantics, not executable	Quick insights, dashboards

Conformance Checking Methods

Method	Pros	Cons	Best Use Case
Token-Based Replay	Fast, intuitive, easy to compute	Less precise, may misrepresent deviations	Quick conformance estimation
Alignment-Based Checking	Very precise, finds optimal matches	Computationally expensive for large logs	Audit scenarios, compliance checking
Log Skeleton	Lightweight, structural conformance	Not as expressive as Petri net alignments	Quick structural validation

Performance Analysis

Technique	Pros	Cons	Best Use Case
Sojourn / throughput times	Easy to interpret, highlights bottlenecks	Needs reliable timestamp data	Detecting slow activities
Time annotations on arcs	Visual enrichment of models	Only as good as the log quality	Identifying bottlenecks in process paths
Case duration analysis	Summarizes case lifetimes	Doesn’t explain internal causes	SLA monitoring

Other Techniques

Method	Pros	Cons	Best Use Case
Trace Variants Analysis	Simple, shows different execution paths	Can explode with many variants	Exploratory analysis
Trace Clustering	Groups similar behaviors	Choice of clustering algorithm impacts results	Finding behavior patterns
Predictive Monitoring (via ML)	Anticipates outcomes, remaining time	Needs feature engineering, external ML models	Predictive SLA, early-warning systems

Key Takeaway:

• If you want robust discovery → use Inductive Miner.
• If you need fast visualization → use DFG Discovery.
• For compliance checks → prefer Alignment-based Conformance.
• For real-life noisy data → Heuristics Miner is strong.