Process Mining

DELIVERY:

Last updated Jun 24, 2025

Discover process model

Introduction

Discovering a process allows you to gain a better understanding of the process structure. Advanced process mining techniques, called Probabilistic Inductive Miner (PIM), automatically identify activities that happen in parallel, are part of a decision, or are part of a more complex loop by analyzing the entire event log. This creates a flexible and realistic representation of how the process actually unfolds in practice, taking into account the different ways that tasks can be performed and their frequency of occurrence. It can be particularly useful in complex environments where processes are not always followed in a strictly defined manner.

You can select Discover process model as the Process mining type when creating a process app. Check out Create new app wizard for more information.

Algorithm

The PIM algorithm works by recursively dividing the set of activities in the event log into smaller subsets until only single activities remain. It then tries to find the best fitting relationship between these subsets each time. These relationships, which define the structure and behavior of the process, are referred to as the process semantics.

PIM works in the following recursive stages:

Directly Follows process model: The Inductive Miner starts with building a Directly Follows process model, which shows how activities follow each other in the event log.
Cut Detection: In the cut detection step, the Inductive Miner divides the process into two parts, such that an operator can describe the relationship between these parts. The operator expresses a "sequence", an "exclusive choice", a "parallelism", or a "looping" relationship between the two halves of the cut.
Log Splitting: In this step, the Inductive Miner applies the split from the cut to the event log, separating the activities of both halves into distinct event sublogs. The algorithm then creates new Directly Follows process models from these sublogs and repeats the process of detecting and splitting cuts recursively. However, before proceeding with the recursion, the algorithm checks for some base cases. For example, if a sublog contains just a single activity, the recursion stops because no further cuts can be discovered.

Cut Detection

The Cut Detection step is the core of the Probabilistic Inductive Miner. PIM calculates behavior scores between 0 and 1 for each pair of activities, indicating the strength of the behavior between them. These scores are derived from the frequencies in the Directly Follows process model data. PIM distinguishes between sequence, exclusive choice, parallelism, and looping behavior.

Next, PIM calculates probabilities for the cut detection. These probabilities are based on the average of activity pair scores between the cut halves. Rather than considering every possible split and picking the best option, PIM directly identifies the cut with the highest probability.

Example

Consider the following event log, and its accompanying Directly follows process model.

<A, D, E>¹² <A, B, C, E>⁵ <A, C, B, E>³

PIM calculates the following sequence behavior scores for every pair of activities. As we always start with A, the scores from A to every other activity are high. As we always end with E, the scores from every activity to E are high.

Sequence scores	A	B	C	D	E
A	-	0.88	0.88	0.92	0.95
B	0	-	0.22	0	0.88
C	0	0	-	0	0.88
D	0	0	0	-	0.92
E	0	0	0	0	0

In the next stage, PIM concludes that both sequence splits (A | B, C, D, E) and (A, B, C, D | E) are equally probable and chooses one of them. The split is applied on the event log, creating two smaller event logs which are recursively analyzed for splits. Finally, PIM identifies a parallel relation between B and C, and a choice relation with D. The following illustration shows the obtained result.