Predictive Process Monitoring Using Graph Neural Networks

The design of the PGTNet model, which predicts remaining time in business processes, is described in detail in this section. Event logs are initially transformed into a graph dataset using this method, in which nodes stand for event types and edges for "directly follows"—direct temporal links. One significant innovation is the inclusion of rich edge characteristics, which provide vital contextual information beyond simple connectivity by encapsulating temporal data (such as durations and timestamps) and system burden (number of active cases).