1. INTRODUCTION

Nowadays, companies try to make their business processes easy to understand and model while growing and opening to the market, so that the process management activity will be done easily and smoothly. Configurable process models (CPM) are useful for large structures that run several activities (Ayora et al., 2013). These models offer the option to regroup multiple variants related to the same business process in one model, which can be customized according to the execution environment properties (Pereira Detro et al., 2020). However, configurable process models present some limits and difficulties such as complexity, redundancy, and misunderstanding (Assy et al., 2014).

Process mining techniques have been very helpful in discovering business process models and configurable process models that represent the real execution of processes (Buijs et al., 2013). Process mining is divided into three types of techniques: discovery, conformance checking, and enhancement. Process model discovery explores event logs to create a clear process model based on the order of activities present in event logs (Ozates et al., 2024). Conformance checking uses the event log to detect any deviation during the process execution. Enhancement aims to improve the process model based on the data present in the event log (Ozates et al., 2024).

In spite of the number of contributions in the process mining field, most of the works have been focusing primarily on the syntactic analysis of process execution traces, which generate multiple issues related to the size and complexity level of the model (Barbieri et al., 2023). And, usually, only business analysts and domain experts can handle analysis regarding the resulting process model, due to the absence of domain knowledge in the process model (Berti et al., 2024). To remedy these limitations, semantics were introduced as a support tool to improve process mining results (Barbieri et al., 2023; Pereira Detro et al., 2020; Rebmann & van der Aa, 2022). For configurable process models, the lack of semantics complicates the task of configuration and impacts on the quality of process model variants (El Faquih et al., 2015; Fang et al., 2023).

As part of this work, we present a new approach to mine configurable process models that are enriched semantically with variability concepts and domain knowledge, starting with a collection of event logs. We aim to capitalize on the semantic analysis of event logs to discover a configurable process model that integrates semantics.

For this purpose, we will use two ontologies, an ontology related to variability concepts and an ontology related to domain knowledge. The main idea is to improve comprehension and quality of the process model discovered. The configurable process model enriched semantically will be simpler to configure. Instead of discovering the model and then checking it against ontologies to validate it, ontologies will be used during the discovery phase to provide additional information, related to variability and process domain, on the process model discovered. This will reduce the cost of conformance and enhancement of these models.

The remainder of this paper is organized as follows. In Section 2, we include a background of the most important concepts. Section 3 gives an overview of the existing work related to this research area. Section 4 describes our objectives and contribution. In Section 5, we present a global overview of our approach. Section 6 provides details about the discovery component. Finally, in Section 7 we conclude the paper with an outlook for perspectives.

2. BACKGROUND

2.1. Configurable Process Models

Configurable process models regroup multiple execution cases of the same business process into one single process model. This type of model presents multiple advantages such as reusability, flexibility, and the ability to fit individual needs without performing any process modeling activities (Fang et al., 2023). Configurable process models have been used in different application areas, especially healthcare (Pereira Detro et al., 2020), industry (La Rosa et al., 2011), and e-learning (Azouzi et al., 2017).

Otherwise, configurable process models can be adapted to the execution context through the configuration decisions made at design time (Assy et al., 2014). The configuration allows users to disable all unnecessary process parts in order to derive the appropriate model (Gottschalk et al., 2009).

Configurable process models are characterized by the support of variability via two key elements: variation points and variants (El Faquih & Fredj, 2017). The variation points represent the decision points that link a fragment of common parts along with a fragment of variable parts in the model, and the variants are the multiple alternatives relating to the variation point. The choice of variants is made based on the rules related to the execution context and variation point. During the configuration process, variant activities can be set as ON (Included), OFF (Excluded), or OPT (Optionally included).

A variety of languages integrate variability in order to represent configurable process models: configurable event-driven process chains (C-EPC) (Rosemann & van der Aalst, 2007), configurable integrated EPC (C-iEPC) (La Rosa et al., 2007), configurable BPMN (C-BPMN) (El Faquih & Fredj, 2017), variant rich BPMN (vrBPMN) (Schnieders & Puhlmann, 2007), and configurable yet another workflow language (C-YAWL) (La Rosa & Gottschalk, 2009). Fig. 1 (La Rosa, 2009) represents an example of the configurable process model and its variants.

 

Fig. 1

Example of the configurable process model. Adapted from La Rosa (2009).

2.3. Semantics in Configurable Process Models

Configurable process models are designed to be changed and customized depending on the execution conditions of the business environment. However, their ability to represent multiple business process cases impacts other properties, and engenders several challenges due to the size of the model, its complexity, and the time plus resources needed to evolve and keep up this model (Detro et al., 2017; El Faquih et al., 2015).

Present research proves that the integration of semantic features into process models can improve model quality and solve multiple misunderstanding issues. Accordingly, the use of semantics in process modeling becomes a necessity, as they bring more clarity and ease during process model manipulation.

The integration of semantics in the configurable process model is inspired from their integration in the business process model. It corresponds to the combination between business process and semantic web technologies. A semantic process model is made up of a range of activities, including additional information related to the functional, behavioral, organizational, operational, and non-functional aspects of these activities. These aspects are described in a formal representation that is machine-readable and machine understandable. The concept of ontologies serves to illustrate these data in process models (Lautenbacher et al., 2009).

An ontology represents a precise and formal specification of a shared conceptualization related to a specific subject area (Gruber, 1993). It defines concepts of a domain and the relationships between them.

Fig. 2 (La Rosa, 2009) illustrates an example of the BPMN configurable process model, annotated semantically using BPM ontology concepts.

 

Fig. 2

Example of the semantic configurable process model. Adapted from La Rosa (2009). DFM, digital film master.

 

The next section is dedicated to the presentation and discussion of the existing works in relation to our study.

3. RELATED WORKS

Over the last few years, the implementation of business process models has been studied, through both theoretical and practical methods. Thus, various studies have identified the business process weakness in dealing with different execution environment conditions. For that reason, configurable process models were introduced with one principal advantage of integrating all the behaviors of all process variants into one model (Gottschalk et al., 2007). This type provides the possibility of extracting the variant models through two techniques, extension and restriction (La Rosa et al., 2017).

In parallel with this, a study by Hepp et al. (2005) showed that business processes need to adopt a machine-readable representation and proposed to combine business process management and semantic web technology under one consolidated technology, semantic business process management, to provide a uniformed representation of business processes on a semantic level.

This evolution has a major impact on process mining results, since the first applications of process mining techniques were based on syntactic analysis, and the models discovered were too complicated and large to be maintained. Many studies have used semantics in different ways to improve the process mining results. Table 2 (Assy et al., 2014; Cairns et al., 2014; Fang et al., 2023; Khodayari Tehraninejad & Khoshnevis, 2022; Nykänen et al., 2015; Pereira Detro et al., 2020; Ramos-Gutiérrez et al., 2021; Rebmann & van der Aa, 2022) presents a comparison of existing works in the field of process model discovery. The comparison study was done using the following criteria:

According to this overview, we can observe that there are two categories of works: the first category (Assy et al., 2014; Khodayari Tehraninejad & Khoshnevis, 2022; Ramos-Gutiérrez et al., 2021) includes works that discover configurable process models or fragments through the application of process mining algorithms on collections of event logs. The resulting model supports variability but lacks semantics. The second category (Cairns et al., 2014; Nykänen et al., 2015; Rebmann & van der Aa, 2022) includes works that discover business process models without variability, but integrate semantics to improve event log data quality or improve the resulting model quality. Some other works (Fang et al., 2023; Pereira Detro et al., 2020) use semantics to derive process variants from an existing configurable process.

Following this comparison, we concluded that most of the existing approaches do not combine variability and semantics in process model discovery. No one of the existing works tries to integrate semantics in configurable process model discovery, despite the importance of this element to improve model quality and facilitate its verification and enhancement. Semantics can be also used as a support tool during process model configuration and analysis.

4. OBJECTIVES & CONTRIBUTION

The objective of this work is to integrate semantics in configurable process model discovery. The use of semantics has produced significant results for business process models. By analogy, we intend to apply the same for CPM, especially since configurable process models are too large and complex and require more effort and resources to analyze, configure, and enhance. Therefore, the usage of semantics will provide rich support during the lifecycle management of CPM.

In this paper, we aim to integrate semantics using ontologies (CPM ontology and domain ontology). The CPM ontology is used to pre-process the event logs and mark variability elements in the resulting merged event log. The domain ontology is used during the process mining application, and related annotations will appear in the resulting configurable process model.

Given that IM is the most used algorithm in process mining practices (Amoury & Bertet, 2023) and can deal with noisy, incomplete traces, we choose to extend it to enrich the resulting configurable process models with semantics.

5. APPROACH OVERVIEW

In our approach, configurable process mining has been realized as a multi-step mechanism, schematized in the below Fig. 3. Below are the three main steps of the proposed approach.

Fig. 3

Semantic configurable process mining approach. CPM, configurable process model.

6. CPM FRAGMENTS DISCOVERY APPROACH

The configurable process model discovery approach is performed on a collection of event logs that is merged and enriched semantically with CPM ontology. The approach proposed is based on multiple steps, as presented in Fig. 3. These steps are described in the next sections.

First, we merge the collection of event logs and annotate the resulting log using the CPM ontology.

We consider that the process we are dealing with is composed of the three types of activities that characterize the configurable process model: i) atomic activity, ii) complex activity, and iii) configurable activity. An atomic activity is a simple activity that performs an atomic task that cannot be divided into sub-activities. A configurable activity includes variability and can be configured during execution. A complex activity consists of more than one atomic activity and configurable activity (Sharma et al., 2014).

In the second step, we will split the merged event log into multiple sub-logs that correspond to CPM fragments, depending on the activity types defined above, and classify them into two categories (with/without variability). Therefore, we will apply mining techniques to discover each CPM fragment separately.

Finally, we will link fragments and integrate activities to obtain the configurable process model.

Input: Collection of event logs+variability specification file+domain ontology+CPM ontology

Output: Configurable process model fragments

6.2. Semantic Annotation of Merged Event Log

The objective of this step is to enrich the merged event log with meta-data related to domain knowledge. For this purpose, we are using a domain ontology that includes a taxonomy classifying process tasks in a hierarchical structure, as shown in Fig. 4 (Paquette, 2010).

 

Fig. 4

Taxonomy of class activity in ontology learning design. Adapted from Paquette (2010). S, SubClassOf; C, Class.

 

Using the knowledge shared about the process, we try to link each event in the merged log to one class of the taxonomy. Through this method, we can determine if the event reports to a complex activity or atomic activity. For example, if our process is composed of the below activities:

CA1: {a1 || a2 || a3}

CA2: {a4 || a5 || a6}

CA3: {a7 || a8}

(CA_i) refers to complex activity, and (ai) refers to atomic activity.

Table 3 demonstrates a fragment of the event log enriched semantically with domain ontology; the event log captures atomic activities executed during the process. For each activity (ai) in event log, we use the column “Semantic Annotation” to reference related complex activity (CA_i).

 

Table 3

Semantically annotated event log

Case ID	Task name	Originator	Timestamp	Semantic annotation
1	a1	User1	20-07-2004 14:00:00	<<CA1>>
1	a2	User1	20-07-2004 14:05:01	<<CA1>>
2	a1	User2	20-07-2004 14:06:00	<<CA1>>
2	a3	User2	20-07-2004 14:20:00	<<CA1>>
2	a4	System	20-07-2004 15:00:00	<<CA2>>
3	a1	User3	20-07-2004 15:01:00	<<CA1>>
1	a4	System	20-07-2004 15:13:00	<<CA2>>
2	a5	User2	20-07-2004 15:27:00	<<CA2>>
3	a3	User3	20-07-2004 16:00:00	<<CA1>>
3	a4	System	20-07-2004 16:09:00	<<CA2>>
2	a7	SYSTEM	20-07-2004 16:35:00	<<CA3>>
2	a8	SYSTEM	20-07-2004 16:38:00	<<CA3>>
1	a6	SYSTEM	20-07-2004 17:00:00	<<CA2>>
1	a7	SYSTEM	20-07-2004 17:14:00	<<CA3>>
1	a8	SYSTEM	20-07-2004 17:22:00	<<CA3>>
3	a5	User3	20-07-2004 18:00:00	<<CA2>>
3	a7	SYSTEM	20-07-2004 18:21:00	<<CA3>>
3	a8	SYSTEM	20-07-2004 18:30:00	<<CA3>>

6.3. Split Merged Event Log

To discover a less complex model, we chose to split the merged event log into multiple sub-logs, where each sub-log corresponds to one complex activity. The idea here is to obtain the sub-logs that will serve as an input for the process mining algorithm to discover configurable process model fragments. Fig. 5 demonstrates the algorithm proposed for this step.

 

Fig. 5

Algorithm for splitting merged event log.

 

This algorithm builds a set of sub-logs based on the merged event log and a list of complex activities present in the log.

For each complex activity, it clones the merged event log and goes through the elimination technique to keep only events related to this complex activity in the copy created. Once we deduced a sub-log for each complex activity of the process, we used the algorithm in Fig. 6 to classify them into two categories: Category 1: sub-logs that contain variability (at least, they contain one variable activity) and Category 2: sub-logs without variability (all activities are common).

 

Fig. 6

Algorithm for classifying sub-logs.

 

Regarding our example, the split event logs are below:

E(CA1) = {{a1 || a2} || {a1 || a3} || {a1 || a3} || {a1 || a2} || {a1 || a3} || {a1 || a2} || {a1 || a3} || {a1 || a2}}

E(CA2) = {{a4 || a5} || {a4 || a6} || {a4 || a5} || {a4 || a5} || {a4 || a5} || {a4 || a6} || {a4 || a6} || {a4 || a6}}

E(CA3) = {{a7 || a8} || {a7 || a8} || {a7 || a8} || {a7 || a8} || {a7 || a8} || {a7 || a8} || {a7 || a8} || {a7 || a8}}

Variable_logs = {E(CA1) || E(CA2)}

Common_logs={E(CA3)}

6.4. Mining CPM Fragments

After classifying the sub-logs into two categories, we will apply the process mining algorithm to discover configurable process model fragments. We are expecting to get for each sub-log a fragment of the configurable process model.

Variable fragments will be discovered from variable sub-logs, and common fragments will be discovered from common sub-logs.

Knowing that the common logs did not contain any variable activity, we will directly apply IM to discover the corresponding fragments, and for variable sub logs, we will apply a customized algorithm Inductive_Miner_Var that includes variability in the process tree discovered.

6.4.1. Mining CPM Common Fragments

For the common fragments we directly apply the basic version of IM (Pohl, 2019), which is based on four main functions: cut detection, log splitting, base cases, and fall through. IM takes as input an event log and builds the appropriate directly-follows graph (dfg). IM looks for a cut in dfg, following this order of priority:

• 1) XOR Cut

• 2) Sequence Cut

• 3) Parallel Cut

• 4) Loop Cut

Depending on the cut found, one of the different split functions below is applied to split the initial log into one or more sub-logs.

• 1) XOR Split

• 2) Sequence Split

• 3) Concurrent Split

• 4) Loop Split

After this step, the algorithm is recursively applied to each sub-log until a base case, i.e. a dfg with a single activity (Leemans, 2017), is detected. In case no base case is found and no more cuts are detected, a fall through is applied (i.e. a process tree is discovered) (Leemans, 2017). Fig. 7 shows the common fragment mined from sub-log E(CA3).

 

Fig. 7

Process tree discovered from sub-log E(CA3).

6.4.2. Mining CPM Variable Fragments

For the variable fragments, we conceived the new algorithm, Inductive_Miner_Var, that represents an extension of the IM supporting variability notions. This algorithm (Fig. 8) is also based on the four main functions of an IM, with further processing to include variability as metadata in the resulting process tree. When a cut is detected, we search for single activities in the cut partitions. Then, we check if this activity is a variant, according to the annotations present in the log; if yes, we mark the activity with the label <<Variant>> and mark the direct relation connected to this activity with one of the labels <<Vp 1..1>>, <<Vp 0..n>>, or <<Vp 0..1>>, depending on the type of the variation point, i.e. Alternatif or Optional or Optional-Alternatif.

 

Fig. 8

Algorithm for mining variable fragments.

 

Fig. 9 illustrates the resulting variable fragment related to the sub-log E(CA1).

 

Fig. 9

Process tree discovered from sub-log E(CA1).

6.5. Consolidation of CPM

The last step is designed for the consolidation of the configurable process model by linking the discovered common and variant fragments. During this step, we annotate common fragments with the label <<Common>> and variable fragments with the label <<VAR>>. Fig. 10 shows the resulting process tree of our example after linking variables and common fragments.

 

Fig. 10

Semantic configurable process tree.

7. CONCLUSION

Seen from an analysis point of view, configurable process models are too large, complex, and difficult to evolve efficiently. Therefore, representing semantics that reference concepts of a formal ontology in those models is much needed.

The semantics of annotations will provide support during the configuration and the extension of the configurable process model. Going further, these labels can be used for the validation of derived variants. In this regard, we focus on discovering configurable process models enriched semantically. Semantics are incorporated using annotations related to variability and domain concepts.

In the present paper, we proposed a novel approach to discover a configurable process model annotated semantically, starting with a collection of event logs and two ontologies: domain ontology and business process ontology. The method consists of discovering common fragments and variable fragments of configurable process models separately, and after that linking them. For variable fragments, we proposed an extension of the IM algorithm to generate variability annotations in the discovered process trees.

The proposed approach shows the theoretical approach related to the creation of annotated models. However, the implementation of the test environment and testing on real logs should be the subject of future research. Moreover, we plan to provide a model evaluation of the proposed method. Also, as future work, we aim to extend this approach to include the discovery of variants/configurations at the same time of the configurable process model discovery through the configuration matrix, and contribute to the enhancement of existing configurable process models using ontologies.