AIMultiple ResearchAIMultiple Research

What are 5 Best Process Mining Algorithms to Consider in 2024?

Process mining algorithms are examples of how machine learning can facilitate process discovery. TThey help clean the required data and generate process models with different strengths and weaknesses. Technical professionals and developers must decide which algorithm to use based on the data and models of the processes they want to automate.

Since it is crucial for developers to select the best algorithm to employ in their process models, in this article, we explain the most common process discovery algorithms.

1. Alpha Miner

Alpha Miner is the first algorithm that bridges the gap between event logs or observed data and the discovery of a process model. Alpha Miner can build process models solely based on event logs by understanding relations and causalities between the steps of processes.

For example, today, several activities are executed in parallel rather than sequentially to minimize the execution time of the process. The alpha miner can be applied to detect parallel activity because it understands start and end events by using double timestamps. As in the figure below, the alpha miner finds out the processes involved in completing A to C as separate traces, then restructures them together in a model by their relations. 

Figure 1: Representation of two traces of event log

To do these, alpha miner generates:

  • A Petri net

The execution of the process starts from the events included in the initial marking and finishes at the events included in the final marking.

Some of the characteristics of the algorithm:

  • It does not support classification: The algorithm does not recognize that one process is the same as another and logs them as independent events.
  • It does not handle noise data well.Noise data is irrelevant or meaningless data that occurs due to data quality issues (e.g., data entry errors or incompleteness). Such data negatively affect the accuracy and simplicity of the process models.
Figure 2: Process model discovered by Alpha miner

2. Heuristic miner

Heuristics Miner is a noise-tolerant algorithm; therefore, it is applied to demonstrate the process behavior in noise data. The Heuristics miner only considers the order of the events within a case. For instance, it creates an event log table with the fields as case id, originator of the activity, time stamp, and activities considered during the mining. The timestamp of activity is used to calculate the ordering of events.

This technique includes three steps:

  1. The construction of the dependency graph
  2. The construction of the input and output expressions for each activity. 
  3. The search for long-distance dependency relations.

Some of the characteristics of the algorithm:

  • Takes frequency of events into account
  • Can only describe events that are either exclusively dependent on each other (AND), or completely independent from one another (XOR)
Figure 3: Process model discovered by Heuristics miner

3. Fuzzy miner

Fuzzy Miner is one of the youngest algorithms suitable for mining less structured processes. These processes exhibit a large amount of unstructured and conflicting behavior (e.g., it turns spaghetti-like models into more concise models.).  The processes may be unstructured and conflicting because they might include information about:

  • Activities related to locations in topology (e.g., towns or road crossings)
  • Precedence relating to the traffic connections (e.g., railways or motorways).
Figure 4: Semantic fuzzy mining: Enhancement of process models and event logs analysis from syntactic to conceptual level

Fuzzy miner applies a variety of techniques, such as removing unimportant edges, clustering highly correlated nodes in to a single node, and removing isolated node clusters. Fuzzy miner cannot be converted to other types of process modeling languages such as Business Process Mining Notations (BPMN) or Business Process Execution Language (BPEL).

4. Inductive Miner

Inductive Miner detects splits [1], which are the conditions or connecting steps between the first and the final stage of the process in the log. There are different types of splits in an event log:

  • Sequential
  • Parallel
  • Concurrent
  • Loop
Figure 5: Discovering Block-Structured Process Models From Event Logs – A Constructive Approach

Once an inductive miner identifies the splits, it recurs on sub-logs until it can find a base case. Inductive miner models give a unique label for each visible transition. These models use hidden transitions specifically for loop splits. a Petri net or process tree can be produced to map the process flow, based on the inductive miner algorithm. 

Figure 6:Process model discovered by Inductive miner

5. Genetic miner

Genetic miner algorithm is derived from biology, specifically from natural selection. It helps deal  with noise and incompleteness in process models. The way it operates is as follows:  

  1. Reading event log
  2. Building initial representation that defines the search space of a genetic algorithm. 
  3. Calculating fitness of each process in this initial representation, also referred as fitness measure. This step  evaluates the quality of a process model point in the search space against an event log. 
  4. Stopping and returning the fittest models
  5. Generating the next model by genetic operators ensures that all points in the search space are defined by the internal representation and  can be reached by running the genetic algorithm. 
gm-fit-ex2.png
Figure 7: Genetic Process Miner

To understand Genetic miner better you can check out the genetic algorithm courses:

How to choose the best algorithm for mining?

It should be compatible

These ML algorithms can be implemented in programming languages like R or Python . Depending on the data type and quality, they can be utilized in all sectors.

It should be interchangeable

In some cases, one database requires more than one miner algorithm because of its structure. There are also situations where several miner algorithms can be used interchangeably to construct similar models. In these cases IT and analyst teams must compare them through the directly-follows graph (DFG) (see, Figure 7) of events and event transitions  to evaluate the models and understand what they capture or omit. Another common practice is to run all the models and measure the fitness, precision, generalization, and simplicity of the models to decide the most applicable model.

Figure 8: Directly-follows Graph (DFG) Example in Healthcare

An event transaction representation of directly-follows graph (DFG)

[1] “Process Mining – Learn Split and Join from Event Log

Further reading

To understand how process mining works, how it differs from automated process discovery, and process discovery tools, check out our in-depth articles:

If you believe your business will benefit from process mining software, feel free to scroll down our data-driven list of process mining software and tools.

Check out our comprehensive and constantly updated process mining case studies list to learn more.

Let us guide you in finding the right vendors:

Find the Right Vendors

Access Cem's 2 decades of B2B tech experience as a tech consultant, enterprise leader, startup entrepreneur & industry analyst. Leverage insights informing top Fortune 500 every month.
Cem Dilmegani
Principal Analyst
Follow on

Hazal Şimşek
Hazal is an industry analyst in AIMultiple. She is experienced in market research, quantitative research and data analytics. She received her master’s degree in Social Sciences from the University of Carlos III of Madrid and her bachelor’s degree in International Relations from Bilkent University.

Next to Read

Comments

Your email address will not be published. All fields are required.

0 Comments