Resumen de Automated discovery of business process simulation models from event logs: a hybrid process mining and deep learning approach
Manuel Alejandro Camargo Chávez
Modern organizations need to constantly adjust their business processes in order to adapt to internal and external changes, such as new competitors, new regulations, changes in customer expectations, or changes in strategic objectives. For example, due to a pandemic, a retailer might experience a 50% increase in their number of online orders while, during the same time, their volume of in-store purchases declines by 30%. To adjust to these changes, the managers may decide to re-deploy employees from the retail stores to the warehouses of the company and the company’s online customer service department. To inform their decisions, the managers need to have an accurate estimate of the impact of the above changes on the delivery and customer service response times. A common approach to make such estimates is to use Business Process Simulation (BPS). BPS refers to the use of computers to explore the dynamics of a business process over time. BPS has long proven to be a useful approach to answer what-if questions in the context of business process redesign. At the same time, the predictions made by BPS models are known to be relatively inaccurate due to the way they are usually applied. Traditionally, domain experts create simulation models manually by using manual data gathering techniques (e.g. interviews, observations, and sampling).
This approach makes the creation of simulation models time-consuming and errorprone. In real-life, business processes tend to be more complex than what domain experts can capture in a manually designed simulation model. Yet, any omission in the simulation model can significantly affect the accuracy and reliability of a simulation. Other limitations of current BPS approaches arise from fundamental assumptions that business process simulation engines make. For example, business process simulation engines assume that human workers work in a robotic (or factory-line) style– meaning that they conduct their work continuously during working hours, without any distractions, without multitasking, and without fatigue. In other words, current business process simulation approaches are not able to capture and reproduce the complexity of human behavior. In this context, this thesis investigates the following overarching question:
How to automatically create accurate business process simulation models based on data extracted from enterprise information systems? Previous research on this question has demonstrated the viability of using a family of techniques for the analysis of business process execution data, known as process mining, to semiautomatically extract BPS models from execution data. Such techniques are fall under the banner of Data-Driven Simulation (DDS). This thesis starts by noting that existing techniques in the field of DDS require manual intervention and fine tuning to produce accurate simulation models. To address this gap, the thesis presents and evaluates a fully automated technique for DDS capable of discovering and fine-tuning BPS models through process mining techniques. The core idea of the technique is to assess the accuracy of a BPS model automatically using a similarity measure that considers both the ordering of activities and their execution times. On this basis, the proposed technique employs a Bayesian optimization algorithm to maximize the similarity between the behavior generated by the BPS model and the behavior observed in the execution data. The thesis, thus, shows that the proposed DDS technique generates models that accurately reflect the ordering of activities. However, the proposed technique often falls short when it comes to predicting the timing of each activity. This phenomenon is due to the assumptions that BPS techniques make about the behavior of resources in the process. To tackle this shortcoming, the thesis combines DDS techniques based on process mining, with generative modeling techniques based on deep learning. In this respect, the thesis makes two contributions. First, it proposes an approach to learn generative deep learning models that are able to produce timestamped sequences of activities (with associated resources) based on historical execution data. Second, it proposes an approach to combine DDS techniques based on process mining, with generative deep learning modeling techniques. The thesis shows that this hybrid approach to learn BPS models leads to simulations that more closely reflect the observed sequences of activities and their timings compared to a DDS technique based purely on process mining and techniques based purely on deep learning.
