Predictive Analytics for Object-Centric Processes: Do Graph Neural Networks Really Help?

The object-centric process paradigm is increasingly gaining popularity in academia and industry. According to this paradigm, the process delineates through the parallel execution of different execution flows, each referring to a different object involved in the process. Object interaction is present, and takes place through bridging events where these parallel executions synchronize and exchange data. However, the complex intricacy of instances of such processes relating to each other via many-to-many associations makes a direct application of predictive process analytics approaches designed for single-id event logs impossible. This paper reports on the experience of comparing the predictions of two techniques based on gradient boosting or the Long Short-Term Memory (LSTM) network against two based on graph neural networks. The four techniques were empirically evaluated on event logs related to two real object-centric processes, and more than 20 different KPI definitions. The results show that graph-based neural networks generally perform worse than techniques based on Gradient Boosting. Considering that graph-based neural networks have training times that are 8-10 times larger, the conclusion is that their use does not seem to be justified.