Enhancing Robotic Process Automation Task Selection: An Integrated Approach Leveraging Process Mining and Feature Extraction

Robotic Process Automation (RPA) is defined as an assembly of tools explicitly designed to mimic human interactions with the user interface (UI) of an information system, thereby automating tasks without necessitating modifications to the underlying infrastructure [5, 8]. This unique ability of RPA to interface with a vast array of disparate, unlinked applications and assimilate them within an existing information system framework has been shown to expedite the development process, yielding a more efficient outcome compared to traditional automation methods.

It has been observed that the deployment of RPA systems significantly reduces human errors. A comparative analysis of RPA systems and human performance, as documented in a recent survey, revealed a notable discrepancy. While human performance demonstrated an accuracy of 90%, RPA systems exhibited an exceptional accuracy rate of 99.9% within auditing systems [7].

The integration of process mining into RPA has been identified as beneficial for process selection. This technique possesses the potential to automate even subprocesses that meet the RPA process selection criteria. In addition, process mining plays a pivotal role in predicting edge cases for transference to human operators. An increasing interest has been observed among vendors to harness the benefits offered by integrating process mining into RPA [8].

Institutional organizations, notably universities, are commonly known to operate a multitude of diverse and unrelated information systems [8]. These systems often span various domains, including learning management, salary administration, and other administrative tasks, and are typically disconnected. However, there remains a necessity for data across these systems, whether inserted, modified, or deleted, to be synchronized. Although these tasks might lack complexity, they are notably repetitive and time-consuming, requiring interaction with multiple user interfaces or systems. The potential for automation of these essential yet under-acknowledged tasks using RPA has been proposed [8].

RPA is achieved by the orchestration of a workflow to execute the processes initially performed by human workers. This is accomplished through the use of modules and functions either designed by vendors or programmed from scratch. The modular nature of RPA, alongside its operation on presentation layers or UIs, provides it with a potential for adoption and allows for agile development. Furthermore, the functional modules of RPA, which can be reused, enable its smooth integration into IT systems [10].

Considering that RPA systems operate on the top layer where human-machine interaction occurs, process mining techniques can be employed strategically to determine the tasks to be automated. This is achieved by analyzing an organization's event log. However, a market demand has been identified for a tool capable of recording and analyzing the logs of interactions between human operators and the UI elements of this top layer, where both RPA and human-machine interactions occur [11].

The utility of RPA systems in the utility sector has been investigated, with a focus on its implementation in the management of electricity billing at Bydgoszcz City Hall, Poland [12]. The results of these investigations have revealed that RPA systems can function as swift and cost-effective tools in the billing management process [12-14].

When appropriately deployed, RPA systems have been recognized for their ability to operate incessantly, thus augmenting efficiency and accelerating processes. These systems also contribute additional value by preserving progress, thereby enhancing accountability, and by their scalability. Furthermore, they are typically more cost-effective compared to traditional software automation systems engineered for process automation [5, 15]. In contrast, other forms of automation software often require modifications to the underlying system, which escalates both the time and cost for implementation [10, 16]. Due to the potential of RPA to reduce operational costs and increase productivity, its incorporation into systems is on the rise among organizations [17-19].

Process mining techniques are principally divided into three categories: the visualization of processes; the comparison of actual processes performed versus the designed process models; and the optimization of process flow [20, 21].

However, a noticeable gap in research has been identified pertaining to the selection of RPA tasks based on real-time data collected during ongoing business processes. Through the utilization of data derived from process mining, tasks can be tailored to the specific requirements of an organization, department, or task type. The combined deployment of RPA and process mining could unearth valuable insights into an organization's business processes. On one hand, the data and insights garnered from process mining can be applied across various stages of RPA [22]; on the other, the logs generated can be analyzed using process mining techniques to enhance the understanding of the business processes executed by the RPA bot [23].

For the selection of tasks for RPA, process mining techniques can facilitate the selection of only those tasks that require and are susceptible to automation using RPA [24, 25]. This method also ensures that decisions are data-driven [26].

A noteworthy partnership between UiPath and Celonis resulted in the addition of functionalities such as the visualization and selection of processes for RPA automation. This collaboration also aided in the development, testing, and deployment of RPA bots [8]. Leopold et al. [27] proposed a method employing supervised learning, a machine learning technique, to identify tasks from their descriptions and subsequently categorize them as fully automated, interactive between machine and human, or manual. Various studies have outlined the essential features required for a task to be successfully automated using RPA [28, 29].

Process mining, a data-driven technique, is employed to obtain insights into an organization's business processes [30]. Typically, data is sourced from the logs of actions performed by employees or machines. This provides valuable insights into how tasks and subtasks are executed, identifies dependencies between activities, detects bottlenecks, and more within the operations of an organization [31, 32].

4.1 Setup

Python 3.9.7 version was used to implement the framework, with pandas 1.3.4, pm4py 2.6.1, matplotlib 3.4.3, and NumPy 1.20.3 libraries.

4.2 Get process mining data

The already created data mining dataset was taken for the implementation purpose. The data consist of the Procurement-to-payment process of a multinational company with 60 subsidiaries, situated in the Netherlands [17].

4.3 Data pre-processing

Table 1. Overview of dataset: Number of events, traces, activities, and workers

Table 2. A sample of an event from the process mining data

Table 3. Overview of dataset: Events each year, traces started and end each year

The initial exploration of the event log helps understand the scope of events, traces, time duration of the dataset and much more. It also reveals the type of activities performed in the given time period. Table 1 outlines the dataset. There are 1,595,923 events with 251,734 traces. Table 2 shows a sample of an event from the dataset. The property concept: name tells all types of activities performed in handling the company’s purchase order. There is a total of 42 activities performed. The dataset is from 1948-01-26 22:59:00+0000 to 2020-04-09 21:59:00+0000 but the number of events from the year 1948 to 2017 and from 2019 to 2020 is negligible as compared to in the year 2018, as shown in the Table 3. Therefore, for the analysis purpose, only the events of the year 2018 are used. Therefore, only the data from 2018 was used.

4.4 Extract activities

For this dataset, all the unique activities are taken for the creation of the matrix. There is a total of 42 activities.

4.5 Get RPA tasks features for automation

For implementing this framework, the features selected with their weights for RPA are shown in the Table 4. Here, the weight of each feature is assumed. In the real-life scenario, the weights should be given according to the importance of a feature needed to be automated.

Table 4. Weights for the selected features

4.6 Create the feature matrix

For the creation of the Feature Matrix, each activity is ranked according to each feature, and then the overall rank of all the activities is calculated by multiplying the rank and weight.

Figure 2 shows how the volume of different activities is comparable to each other.

In the dataset, there were two types of uses, ‘Batch’ and ‘Users’ as define in the property org: resource as shown in the sample event from Table 2. The Users are human employees, from this, it can be found out which activity is manual, and which is automated. The blank value is not counted. Figure 3 shows the different manual and automated activities.

Error-prone activities are checked by finding the repeated activities in each trace. Figure 4 shows the number of times an activity is repeated in each trace.

Finally using these features, we can get the process for RPA. One limitation of this proposed framework is that the features and their weights must be independently decided by the organisation according to their need for RPA task selection. The efficiency of the framework highly depends upon the list of features with their respective weights for the creation of feature matrix. While the framework is able to extract the activities according to the features given by the organisation, the independently identification of the features with their weights can be difficult for the organisation. This can give rise to the need of an expert for feature identification, increasing the cost for RPA implementation in the process. Future research can include a method to automatically extract the feature with their weights from process mining data, which can reduce the manual work.

2.jpg

Figure 2. Volume of each activity in the process mining dataset

3.jpg

Figure 3. Manual vs. automated activities

4.jpg

Figure 4. Error prone activities

The rank for the activity, “Change Quantity” is:

= Feature Rank of Volume * Weight of Volume +

Feature Rank of Manual Activities * Weight of Manual Activities +

Feature Rank of Error Prone Activities * Weight of Error Prone Activities

= 10 * 3 + 8 * 5 + 6 * 4

= 30 + 40 + 24

= 94

And the rank for the activity, “Delete Purchase Order” is:

= Feature Rank of Volume * Weight of Volume +

Feature Rank of Manual Activities * Weight of Manual Activities +

Feature Rank of Error Prone Activities * Weight of Error Prone Activities

= 12 * 3 + 10 * 5 + 21 * 4

= 36 + 50 + 84

= 170