PhD Defense: 'Deep Learning Models for Predictive Monitoring of Business Processes'

Process mining is a discipline that bridges the gap between traditional model-based analysis techniques and data-centric approaches. This discipline uses event logs to extract knowledge about the actual performance of business processes. Process mining has evolved rapidly, offering powerful tools for analysis, discovery, optimization and improvement of complex business processes. One challenge in this area is predictive monitoring, which focuses on predicting, from different points of view, how a running case will develop. 

Some relevant predictions are the prediction of the next activity or the sequence of next activities. Traditional predictive monitoring methods usually encounter many difficulties due to the dynamic and complex nature of real-world business processes, which leads to less accurate and robust predictions.  In this dissertation, predictive monitoring in process mining is improved by using advanced deep learning techniques. By integrating graph neural networks with recurrent neural networks, direct learning about the process model is enabled while also allowing the information contained in the sequence of events to be taken into account. Thus, two neural models are introduced: the first one predicts the next activity in a business process while the second one predicts the remaining sequence of activities until the end of the running case. For the second problem, a novel model based on deep reinforcement learning is also proposed such that the optimal activity selection strategy is learned during network training. All models have been rigorously validated using real event logs using a novel evaluation methodology in the field that facilitates robust and fair comparison between different predictive monitoring approaches.

Supervisors: Manuel Lama Penín and Juan Carlos Vidal Aguiar