All Times ET
Keywords: GAN, Process, Prediction, Conditional, Wasserstein
Predictive business process monitoring (PBPM) aims at predicting the future of running process instances, be it for the next event or the remaining sequence of events (suffix). An event is characterized at least by a triplet made of a unit identifier, an activity the unit can go through, and the time of activity execution. For suffix predictions, generative networks (GAN) have proven to be the most efficient (Taymouri and La Rosa, 2020). To improve further on the model, we changed it to a Wasserstein GAN (Arjovsky et al., 2017) to stabilize and ease its training phase. However, our main contribution, comes with feature engineering: predicting an instance’s suffix from activities and dates is restrictive, particularly for remaining time predictions. External, unobserved elements impact the progress of a process instance, and covariables are hardly ever taken into account in current models. In the absence of covariables, the immediate data we can generate from process event logs is the crowding of each activity at any given time, as this will be an obvious cause of delays and reroutings of running process instances. Since this data is not considered by current models, they predict suffixes only by taking an instance’s available past and current activities, as well as corresponding dates, which does not contain information of process crowding and, thus, does not allow the model to estimate if a unit will contain delays or reroutings based on the process’ state as it evolves through time. We, therefore, propose a feature engineering method to extract process crowding data and predict suffixes conditionally to such variables (Mirza and Osindero, 2014) by turning the model into a conditional WGAN. We will then show its efficiency at proposing improved time predictions and potential alternative, more accurate routes for process instances.