Deep Learning Models for Predictive Monitoring of Business Processes

Process mining is a discipline that bridges the gap between traditional model-based process analysis and data-centric approaches. It leverages event logs to extract insights about the actual performance of business processes. This discipline has evolved rapidly, offering powerful tools for the analysis, optimization, and enhancement of complex business processes. A major challenge within process mining is predictive monitoring, which focuses on predicting future aspects of ongoing cases, such as the next activity or the remaining sequence of activities. Traditional methods often struggle with the dynamic and complex nature of real-world processes, leading to less accurate or robust predictions. In this thesis, we enhance predictive monitoring in process mining through the use of advanced deep-learning techniques. By integrating Graph Neural Networks with Recurrent Neural Networks, we learn directly from the process model while also considering event sequences. We introduce two neural models: the first aims to predict the next activity in a business process, while the second forecasts the remaining sequence of activities until the case finishes. For the latter problem, a new Reinforcement Learning model is also proposed to dynamically learn optimal activity selection strategies during training. All models are rigorously validated using real-world event logs under a novel evaluation methodology to facilitate robust and fair comparisons between different predictive monitoring approaches.

keywords: Process mining, Deep Learning, Predictive Business Process monitoring, Graph Neural Networks, Reinforcement Learning