Activity Prediction of Business Process Instances using Deep Learning Techniques

The ability to predict the next activity of an ongoing case is becoming increasingly important in today’s businesses. Processes need to be monitored in real-life time in order to predict the remaining time of an open case, or also to be able to detect and prevent anomalies before they have a chance to impact the performances. Moreover, financial regulations and laws are changing, requiring companies' processes to be increasingly transparent. Process mining, supported by deep learning techniques, can improve the results of internal audit activities. The task of predicting the next activity can be used in this context to point out traces at risk that need to be monitored. In this way, the business is aware of the situation and, if possible, can take resolution actions in time. In recent years, this problem has been tackled using deep learning techniques, such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) neural networks, achieving consistent results. The first contribution of this thesis consists of a generation of a real-life process mining dataset based on the Purchase-to-Pay (P2P) process. The SAP tables structure is taken into account since it is the most popular management software in today's companies. We exploit the simulated dataset to explore modeling techniques and to define the type and the quantity of anomalies. The second contribution of the thesis is an investigation of LSTM neural networks architectures that exploit information from both temporal data and static features, applied to the previously generated dataset. The neural networks are then used to predict future events characteristics of running traces. Finally, real-life application of the results are discussed and future work proposals are presented.