Sliding and Adaptive Windows to Improve Change Mining in Process Variability

Abstract

A configurable process Change Mining approach can detect changes from a collection of event logs and provide details on the unexpected behavior of all process variants of a configurable process. The strength of Change Mining lies in its ability to serve both conformance checking and enhancement purposes; users can simultaneously detect changes and ensure process conformance using a single, integrated framework. In prior research, a configurable process Change Mining algorithm has been introduced. Combined with our proposed preprocessing and change log generation methods, this algorithm forms a complete framework for detecting and recording changes in a collection of event logs. Testing the framework on synthetic data revealed limitations in detecting changes in different types of variable fragments. Consequently, it is recommended that the preprocessing approach be enhanced by applying a filtering algorithm based on sliding and adaptive windows. Our improved approach has been tested on various types of variable fragments to demonstrate its efficacy in enhancing Change Mining performance.

1. Introduction

Changes impact all domains, affecting the behaviors of systems, models, and concepts throughout their life cycles. Business processes are no exception; they, too, face similar situations, and changes affecting them must be addressed by business process managers to align the processes with new circumstances.

Over time, many approaches have been proposed to deal with changes in business processes. This began with business process engineering, initially with the WIDE [1] and ADEPT [2] frameworks, which adopt dynamic management of business processes. In the same vein, the change patterns proposed by [3] assist in managing changes by providing patterns that guide managers in handling these changes. Recently, with the emergence of new technologies and intelligent solutions, managing business processes has become more advanced and intelligent, particularly with the introduction of process mining.

Process mining involves a set of solutions that utilize intelligent approaches developed from data science to detect business process models, verify conformity, perform conformance checking, and conduct Change Mining, among other approaches [4].

In this paper, we focus on Change Mining, an area enriched with many intelligent approaches that assist business process managers in managing changes, as highlighted in various studies [5,6,7,8]. However, addressing each business process model independently within organizations that employ multiple versions of the same process is both time-consuming and resource-intensive. To address this, configurable process models are employed. These models consolidate all variants of a process used within an organization into a single reference model, which not only encapsulates all possible variants but also facilitates centralized management of these variants through the reference model.

Therefore, to effectively manage changes in configurable process models, particularly those affecting the variable parts, we utilize a Change Mining approach that requires specific data sources aligned with the needs of Change Mining in such models. A common element among all process mining approaches is their reliance on analyzing event logs. An event log is a file that documents the behavior of the business process; it records detailed information about each event as it occurs.

However, while process mining practitioners often assume that event logs are universally available, these logs require preprocessing, a crucial first step. This step is the initial component of our Change Mining framework and is particularly important when analyzing a collection of event logs. The next component is a Change Mining algorithm specifically adapted to the context of variability. Finally, a change log model is used to record the detected changes.

In this paper, we highlight our previously proposed Change Mining framework for configurable processes, which integrates preprocessing, a Change Mining algorithm, and the generation of change logs. We also suggest an update to our preprocessing approaches, demonstrating improved results over the initial methods proposed in our earlier work. Furthermore, we test the framework on synthetic data to better showcase its ability to detect changes in a collection of event logs associated with process variants from the same family.

The remainder of this paper is organized as follows: Section 2 introduces important concepts relevant to our work and reviews related literature. Section 3 describes the Change Mining framework. Section 4 details the new filtering algorithm. Section 5 discusses the results obtained from applying our new approach to synthetic data. Finally, the last section concludes the paper.

2. Concepts and Definitions

In this section, we present the most important concepts related to our context. We begin by defining the configurable process, followed by an explanation of changes within configurable processes. Finally, we define Change Mining.

2.1. Configurable Process

To define a configurable process, it is necessary to first define a business process.

Business Process: A business process is a collection of interconnected operations or activities that are organized and structured to accomplish a specific business goal [9]. It describes how work should be conducted, who is responsible for what, and the desired outcomes. The purpose of a business process is to optimize and streamline operations, increase efficiency, and improve overall performance [10]. Various modeling languages have been suggested to display all the elements of a business process. Examples of business process modeling languages include Event-driven Process Chain (EPC) [11], Unified Modeling Language (UML) [12], and Business Process Model and Notation (BPMN) [13].

Organizations often use several variations of the same business process to achieve the same goal under different conditions, resulting in multiple process variants. To manage these variants efficiently, the concept of a configurable process has been proposed, which is based on the principle of variability.

Variability: The concept of variability is introduced in different contexts. In product line engineering, variability allows for product differentiation and diversification, referring to the ability of an artifact to be configured, customized, extended, or modified for use in a specific context [14]. In component engineering, the variability of a component includes its structural variability, functionalities, and interaction contracts [15]. In the context of a business process, process variability refers to the diversity of manufacturing processes used to produce product variants within a product family [16]. Managing variability in a business process helps ensure consistency, reduce errors, and improve overall quality and performance. By introducing reuse in business processes, variability can be managed effectively, and configurable process models exemplify the use of variability in business processes.

Configurable Process: A configurable process is a generic conceptual model that formalizes recommended practices for reference models, aiming to capture reusable practices. It is motivated by the principle of “design by reuse” [17], and it is a model that can be modified or adjusted to meet specific needs or requirements [18]. This approach allows for the customization of process models to fit the unique requirements of a particular need or project. The configurability of the process makes it more flexible and adaptable to changing requirements, making it a valuable tool for process improvement and optimization. Compared to a traditional business process, a configurable process includes the following features:

• The variation points: Elements likely to be changed.
• Variations: Options available for each variation point.
• Configuration: The process of creating a configurable process variant by selecting an option for each variation point. The options include choosing a specific variant, excluding a variant, or making a selection based on certain conditions.

Thus, when dealing with configurable business processes, it is important to use specific modeling languages such as C-EPC [19] and C-BPMN [20], which can describe variation points, variations, and configuration choices. Additionally, it is possible to use a variability file [21], which lists variation points and their corresponding variants in a structured file.

Managing business processes or configurable processes involves conception, implementation, monitoring, and updating. All these tasks are performed within the context of Business Process Management (BPM). BPM aims to provide all the necessary approaches to efficiently manage business processes and configurable processes, with one of its important focuses being the management of changes. However, before presenting approaches for managing changes in configurable processes, it is essential to highlight the nature of changes in configurable processes.

2.2. Changes of Configurable Process

The ability to modify a configurable process increases its flexibility and adaptability to changes, which is critical for organizations that must remain competitive and responsive to a rapidly changing environment. However, managing these modifications presents several challenges, especially if changes are not documented. Such changes can be driven by evolving business needs, technological advancements, or the adoption of new regulations or standards. Changes in business processes can include modifications to:

• Control flow: the sequence of tasks or activities that make up a process;
• Tasks: distinct activities or steps within a process;
• Roles and responsibilities: the distribution of tasks among specific individuals or groups;
• Inputs and outputs: the data or information that flows into and out of a process;
• Performance metrics: the measures used to evaluate the success of a process;
• Policies and procedures: the rules and guidelines that govern a process.

All these modifications are frequently the result of managerial decisions. However, other types of changes can occur randomly, such as concept drift [22], also known as process drift [23]. These are unexpected behaviors of the process that cause it to be executed in a different manner.

However, in a configurable process, changes also affect the variable elements, which include (i) variation points, (ii) variants, and (iii) configuration constraints. To manage these changes, many approaches have been proposed. In the next section, we present these approaches.

2.3. Change Mining

By the early 2000s, many BPM practitioners began using process mining [4] to manage their business processes. This technique provides various intelligent approaches that can, from stored data in event logs, discover models, verify conformity, and enhance business process models. Change Mining is a subset of process mining that focuses on analyzing data to detect changes. The data used for Change Mining can include models, event logs, or change logs. However, the most commonly used approaches are based on event logs because change logs are not always available [24], and models are often too complex to analyze.

To achieve the goals of Change Mining, many approaches have been proposed. These include model-to-log alignment [8,25,26], graph-based analysis [24,27,28], clustering-based approaches [5,29,30,31], and windowing and statistical analysis [6,32,33,34]. However, all these approaches are typically applied in the context of a single business process model, and few studies have addressed changes in configurable process models using Change Mining.

Therefore, we propose a comprehensive framework for managing changes in configurable process models. In the next section, we present the proposed framework and its various components.

3. Framework of Configurable Process Change Mining

In this section, we present our comprehensive framework for detecting and managing changes in configurable process models. This framework integrates multiple components, including preprocessing, Change Mining algorithms, and the generation of change logs [35]. The following subsections provide an in-depth look at each component and their roles in the overall framework.

3.1. Overview of Configurable Process Change Mining Framework

In order to automatically detect and record the changes that a configurable process encounters during its life cycle, particularly during its execution phase, we propose a framework for mining changes in a collection of event logs. Our change detection framework enables the identification of instances that contain changes due to both voluntary and involuntary actions affecting the process. As illustrated in Figure 1, our framework consists of three components:

• First Component: Preprocessing. This component builds the event log of variable fragments. It contains two modules: the first module merges event logs related to a family of process variants, and the second module filters events related to variable fragments of the merged event log.
• Second Component: Change Mining Algorithm. This component involves the Change Mining algorithm, which is based on an extension of the STAGGER algorithm [36] dedicated to detecting concept drift.
• Third Component: Generating the Change Log. This component is responsible for generating the change log of variable fragments for the configurable process.

3.2. First Component: Preprocessing

The data used as input for our Change Mining framework are a collection of event logs. However, this collection requires additional preprocessing before it can be used effectively for Change Mining. For this purpose, we propose two approaches: the first is a merging approach, and the second is a filtering approach, as presented in Figure 2.

3.2.1. Merging Approach

The merging approach aims to combine all the event logs in the collection into a single event log, allowing us to use one data source for Change Mining [37]. The resulting merged event log contains all events related to the execution of all process variants, with each event linked to its respective process variant by a unique identifier that specifies the source process variant.

Figure 1. Framework of Change Mining in a collection of event logs [35].

Figure 1. Framework of Change Mining in a collection of event logs [35].

Figure 2. Preprocessing approaches [38].

Figure 2. Preprocessing approaches [38].

However, due to the large number of events in each event log, the resulting merged event log is enormous. Therefore, we need to filter the merged event log to obtain a smaller, more manageable event log. Since we are primarily concerned with the variable fragments of the configurable process, we need to focus on only the parts of the event logs related to these variable elements.

3.2.2. Filtering Approach

To filter events related to variable fragments, we first identify the variable fragments of the configurable process. These are defined as sub-parts of the configurable process that contain at least one variation point [39]. In our work, we further refine this definition by specifying that the variation point should be in the middle of the variable fragment. Elements directly connected to the variation point are termed “previous” and “next”, where “previous” refers to the element preceding the variation point and “next” refers to the element following the variation point.

All events related to elements of variable fragments need to be analyzed. Therefore, we propose an approach that selects events related to variable fragments from the merged event log. The filtering algorithm, as presented in [37], compares the elements of each event with the list of elements in the variability file. When an event related to a specific variable fragment is identified, all events related to the same fragment are selected. Our goal is to capture all events related to the elements of each variable fragment of the configurable process.

The selected events related to variable fragments are stored in a new event log, which we refer to as the event log of variable fragments.

This event log of variable fragments serves as the input for the next component of the framework, which is the Change Mining approach.

3.3. Second Component: Change Mining Approach

The Change Mining approach proposed for detecting changes from the event log of variable fragments is based on an incremental learning algorithm inspired by a concept drift detection method called STAGGER [36]. Our algorithm, as presented in our previous work [40], consists of seven steps that repeat for each trace in the event log of variable fragments. The goal is to create a list, called a grid of detected changes, containing all events related to elements not present in the variability file. The steps are as follows:

• Classification: In this step, all possible fragments are created by generating all possible combinations that can form variable fragments. The data used for this purpose is solely the variability file. This step is performed only once, and the algorithm does not loop through it. The loop starts from the second step.
• Initialization: This step involves selecting the fragment of events for analysis.
• Projection: In this step, events are compared to elements on the list of variable fragments.
• Evaluation: This step allows us to determine if the fragment being analyzed is subject to change and to identify the element that changes.
• Aggregation: During this step, an identifier is assigned to each variable fragment that changes.
• Storage: In this step, events with detected changes are saved in the change grid.

For a more detailed explanation of these steps, please refer to our previous work [40]. Once the change grid is obtained, it is important to transform this grid into an easy-to-read document. Therefore, we propose converting the change grid into a change log. This change log transformation constitutes the third component of our framework.

3.4. Third Component: Generating Change Log of Variable Fragments

We propose presenting the discovered changes in a change log of variable fragments, which will organize the changes and make them easier to read and analyze. The following details will be listed in the change log:

• Change date and time;
• Trace identifier;
• Identifier of the process variant;
• Variation point;
• Type of change;
• Changed item name;
• The name of the new item as a result of the change.

All these details are added to each change event, and for simplicity, we use an XML format to represent this information in an organized manner. Figure 3 provides an example of a sub-part of the obtained change log of variable fragments.

This proposed framework can detect many types of changes caused by a variety of sources. Additionally, all detected changes can be stored in a manner that makes them simple to comprehend and analyze. However, in some cases, the framework is unable to detect certain changes, especially when dealing with large models that have variation points at the beginning or end of the model. We identified that fragments at the beginning or end are not selected by the filtering algorithm. To address this issue, we propose a new filtering algorithm based on a sliding and adaptive window, which can detect smaller fragments. In the next section, we present our new filtering algorithm.

4. The Sliding and Adaptive Window for a Collection of Preprocessing Event Logs

Windowing is a type of fragmentation that allows us to reduce the quantity of data to be analyzed while simultaneously building comparable fragments to determine the non-conformities between different executions of the process. In the next subsection, we present various types of windowing used in Change Mining.

4.1. Types of Windowing Approaches Used in Change Mining

Researchers have adapted and developed different types of windowing approaches for use in the context of Change Mining.

• Fixed Window: The size and properties of the window used for fragment selection are fixed and specified at the beginning of the analysis [39].
• Time Window: This change detection method uses one window to show past events and another window to show recent events. It then compares the distributions over these two windows using statistical hypothesis testing [41].
• Adaptive and Sliding Windows: A window of fixed size determines the instances of incoming data used for learning. Depending on the case, this window will be adapted to determine the new characteristics of the window [42].

In the context of a collection of event logs, we use the windowing approach to select events related to elements of variable fragments from the merged event log. In the next subsection, we explain how we adapt the windowing approach to meet the needs of Change Mining in configurable processes.

4.2. The Sliding and Adaptive Windows to Improve Configurable Process Change Mining

Since variable fragments are sub-parts of the configurable process, the appropriate type of window to use in this case is the adaptive window, which can adjust to the size of each variable fragment. Additionally, we need a sliding window to move through the events in the merged event log.

The sliding and adaptive window, in this case, will select a block of events in each loop, composed of either two or three events depending on the position of the window. If the window is at the beginning of a trace, it will select a block of only two events, and the same applies at the end of a trace. However, in the middle of a trace, all blocks will be constrained to three events. Additionally, the construction process of these blocks must follow specific rules to achieve the desired results:

• Uniform block: All elements of the block must have the same trace identifier (TRACE-ID) and the same process variant identifier.
• Parallel Gateway Handling: If a parallel gateway, also known as an AND gateway, is the first element of the block, we loop through the events of the AND gateway until we reach its last element. This last element will then be the first element of the new block.
• End-of-Block with an AND Gateway: If an element of an AND gateway is at the end of a block, we keep the block as is. For the selection of the next block, we advance to the last element of the AND gateway, making it the first element of the new block.

After looping through all events in the merged event log and applying these rules to construct the blocks, we will have a list of blocks. From this list, we will identify the blocks that represent variable fragments.

A block is considered to be a block of events of a variable fragment if it satisfies the following conditions:

• One of the block elements exists in the variability file.
• The position of this element within the block corresponds to the same position predefined in the variability file. Specifically:
  • If the element corresponds to a variant, it must be in the middle, as this is the position considered for building variable fragments of the configurable process.
  • If the element is predefined in the file as “next”, it must be at the end of the block.
  • If the element is predefined as “previous”, it must be at the beginning of the block.

To illustrate our filtering algorithm, Figure 4 presents an example of a configurable process model with a single variation point. We generate three process variants and demonstrate how our algorithm processes events, block by block, to select events related to variable fragments.

The algorithm starts by creating the first block with the first two events, E1 and E2. These two events are related to elements of an AND gateway. According to our block creation rules, we continue the selection until the last element of the AND gateway, which, in our example, is related to event E3. The following block will, therefore, be E3-E4-E5. We then compare the elements of the block with the elements of variable fragments in the configurable process model. The variable elements can be read from the configurable process model or the variability file.

This comparison starts with the first element related to the first event in the block. This first element corresponds to an element that exists in the variability file, and it is identified as the previous element. Consequently, the comparison stops here, and the block will be recorded in the event log of variable fragments with the identifier “Id of variable fragment = 1” for the three events in the block.

Next, we move to the block composed of events E4-E5-E6 and perform the same comparison. The element of the first event exists in the variability file, but its position is different. The element of the third event does not exist in the variability file. Consequently, we skip this block and move on to the next one, which is E5-E6-E7. The TRACE-IDs of these three events are not identical, indicating that we have reached the end of the first trace.

Therefore, for the last block, we keep only two elements, E5 and E6. We then perform a new comparison, which results in ignoring the block since the two events do not correspond to elements of variable fragments.

The selection of the next block starts with a new trace, so the block will contain only two events, E7 and E6. The algorithm will repeat the same process of building blocks and selecting events until it reaches the end of the merged event log.

By using this new filtering approach, it is possible to select all events of variable fragments, even if the variable fragments exist at the beginning or at the end of the configurable process model. This will help improve the results of Change Mining.

In the next section, we demonstrate this capability by experimenting with our new algorithm and comparing its performance with the old filtering algorithm. Additionally, we will test the ability of our Change Mining framework to detect changes in a collection of event logs.

5. Experience

In this section, we detail the methodology and results of our experiments designed to test the effectiveness of our proposed algorithms. We begin by describing the synthetic data used in our tests, followed by a comprehensive analysis of the experimental results. This analysis includes a comparison between the old and new filtering algorithms, highlighting the improvements achieved by the new approach.

5.1. Data Used for the Test

To test our solutions, we implemented all our proposed approaches as functions within our tool, the Random Configurable Process Model and Log Generator (rCPMLG) [37]. We applied both the old and new filtering algorithms to different configurable process models. The data used for the test are synthetic data generated by rCPMLG, which randomly create configurable process models and their process variants and simulates the execution of these variants, storing the results in event logs. The strength of this tool lies in its configurability, allowing us to generate models according to specific needs, including the number of variation points and variants for each variation point. Additionally, it can generate event logs with changes, which is crucial for our tests.

For our simulation, we conducted ten experiments on ten different collections of event logs. Each pair of collections came from a configurable process model with the same number of variation points and variants, and we used the same number of process variants. However, the position of the variation points differed. Our goal was to test the ability of our algorithm to detect variable fragments in different positions.

Table 1. Data used for the experience.

N°	NVP 1	Nbr Variants 2	Types 3	N logs IN C 4
1	3	[3,2,3]	[N.N]	6
2	3	[3,2,3]	[Y.Y]	6
3	4	[3,4,3,4]	[N.N]	8
4	4	[3,4,3,4]	[Y.N]	8
5	5	[3,4,5,3,4]	[N.N]	10
6	5	[3,4,5,3,4]	[N.Y]	10
7	6	[3,4,5,3,4,5]	[N.N]	12
8	6	[3,4,5,3,4,5]	[Y.Y]	12
9	7	[3,4,5,3,4,5,3]	[N.N]	14
10	7	[3,4,5,3,4,5,3]	[Y.Y]	14

¹ Number of variable points in the configurable process. ² Number of variants for each variation point. ³ The presence or absence of a variation point at the start or at the end of the configurable process. ⁴ Number of event logs in the collection.

presents the properties of the collections of event logs used in each experiment, indicating the presence or absence of a variation point at the start or end of the configurable process. In each experiment, we generated event logs containing random changes.

These data, presented as a collection of event logs, will serve as the input for our framework. In the next section, we present the results of the experiments.

5.2. Experiences and Results

In this section, we present the results of our experiments on different collections of event logs, generated from various configurable process models.

We applied two approaches to each collection, resulting in two experiments for each collection. The first experiment used the event log of variable fragments obtained with the old version of the filtering algorithm. The second experiment used the event log generated by the new filtering algorithm based on sliding and adaptive windows.

Since synthetic data has been used and random changes applied using our tool, it is possible to know the number of changes applied, although the value and position are not predefined. The detection percentage can be calculated by dividing the number of detected changes (DCs) by the number of applied changes. To verify the algorithm’s ability to detect variable fragments, the number of detected fragments (DFs) has been compared with the number of possible variable fragments. This number has been calculated by multiplying the number of traces in each event log by the number of variation points and the number of process variants in the collection.

Table 2. Results of the experience.

			First Experience 1			Second Experience 2
N°	Variable Fragments	Applied Changes	DF 3	DC 4	%	DF 3	DC 4	%
1	1800	153	1797	149	97%	1800	149	97%
2	1800	493	1184	239	49%	1799	492	99%
3	3200	546	3198	544	100%	3201	544	100%
4	3200	186	3198	184	99%	3200	186	100%
5	5000	629	4998	629	100%	4950	501	80%
6	5000	600	3999	404	67%	4923	546	91%
7	7200	630	7200	607	96%	7200	558	89%
8	7200	464	5984	311	67%	7156	461	99%
9	9800	1129	9800	1125	99.6%	9800	1125	99.6%
10	9800	439	8400	271	61.7%	9793	431	98%

¹ First experience is made by using the first algorithm based on comparing events in the event log consequently. ² Second experience is made by using our new algorithm based on sliding and adaptive windows. ³ DF (detected fragment): This refers to the number of variable fragments identified by the filtering algorithm. ⁴ DC (detected change): These refer to the number of fragments with changes identified by the Change Mining algorithm.

summarizes the results of our experiments with different configurable process models. We observe from the results that both approaches give approximately acceptable results when the model does not have a variation point at the beginning or end of the configurable process. However, if the configurable process has a variation point at the beginning or end of the model, the algorithm does not detect all the variable fragments. Consequently, the number of detected variable fragments is fewer than the number that should be detected.

5.2.1. Illustrative Graphs

To better visualize the results, we created two graphs that show the ability of our algorithms to detect changes in two scenarios. The first scenario involves a configurable process model without a variation point at the beginning or end (see Figure 5). The second scenario involves a configurable process model with a variation point at the beginning or end (see Figure 6).

5.2.2. Detailed Analysis

In the first graph (see Figure 5), we observe that the two lines describing the detection ability are in the same position.

In the second graph (see Figure 6), we observe that the two lines are significantly different. The first algorithm does not detect all changes, whereas our new algorithm is able to detect nearly all the applied changes.

From the same graphs, we also observe that the size of the data used for the configurable process model does not affect the results of our solutions.

5.2.3. Summary

Overall, our experiments demonstrate that the new filtering algorithm significantly improves the detection of changes in configurable process models, particularly in cases with variation points at the beginning or end of the model. This enhancement supports the viability of our approach for more accurate and efficient Change Mining.

6. Conclusion

In this paper, a framework for detecting and managing changes in configurable process models has been expanded upon, building on our previous work. This framework integrates preprocessing, an advanced Change Mining algorithm, and the generation of change logs to effectively handle the complexity and variability inherent in configurable processes.

A novel filtering algorithm based on sliding and adaptive windows has been introduced to address the limitations of traditional methods, particularly in scenarios where variation points are located at the beginning or end of the process model. The experiments conducted using synthetic data generated by the Random Configurable Process Model and Log Generator (rCPMLG) have demonstrated that the new filtering algorithm significantly improves the detection of variable fragments, ensuring more accurate and reliable Change Mining results.

Through a series of tests on various collections of event logs, it has been shown that our approach is robust and adaptable, capable of handling different configurations and sizes of process models without compromising performance. The analysis has further validated the effectiveness of our algorithm, highlighting its superiority over previous methods.

However, when changes affect all elements of the same variable fragment, it becomes impossible to detect the events of variable fragments, and thus these changes cannot be detected. In our future work, an update to our Change Mining algorithm will be proposed to detect more complex changes. Additionally, the application of this framework in real-world scenarios could be explored, extending its capabilities to handle more complex process configurations.

The contributions of this work provide a valuable tool for business process management, enabling organizations to maintain flexibility and adaptability in their processes while efficiently detecting and managing changes. Overall, our framework represents a significant advancement in the field of Change Mining, offering a practical and effective solution for managing configurable process models in dynamic and evolving environments.