Suffix prediction of business processes forecasts the remaining sequence of events until process completion. Current approaches focus on predicting the most likely suffix, representing a single scenario. However, when the future course of a process is highly uncertain and variable, a single scenario may have limited predictive value. To address this limitation, we propose probabilistic suffix prediction, a novel approach that returns a set of sampled suffixes. The method is based on an uncertainty-aware encoder-decoder LSTM combined with a Monte Carlo suffix sampling algorithm. We capture epistemic uncertainty via MC dropout and aleatoric uncertainty as learned loss attenuation. Comparisons with two other uncertainty-aware PPM approaches across four datasets demonstrate that our probabilistic suffix prediction approach achieves reasonable predictive performance while it allows estimating prediction intervals for multiple objectives within a single model.

Allocating resources to process tasks during runtime (online) is hard. A solution method for such allocation is required to be computationally efficient while being subjected to uncertainties such as resources suddenly becoming (un)available. Resource allocation problems where task processing times differ across resources can be formalized as an assignment or parallel machines scheduling problem. This work presents adaptations to both problem formulations to address resource (un)availabilities. These adaptations require a prediction model to estimate the processing time of a task for all of its authorized resources. We evaluate and compare the proposed adaptations with existing allocation approaches on two process simulation models created from an artificial and a real-life event log. Our results show that both approaches can outperform traditional allocation strategies, such as the shortest queue, random, round-robin, and batch-allocation approaches.