Business process analysis and improvement approaches typically require accurate models of activity processing or preceding waiting times. Data-driven methods derive these from historical event log data, but typically fit static probability distributions that overlook contextual information or train regression models to produce point estimates, disregarding potential uncertainties. To address these limitations, we present the Pro3Log approach that enables Probabilistic learning of Processing and waiting times of activities based on Process event logs. Probabilistic learning is a machine learning approach that enables the learning of dynamic probability distributions, which can model uncertainties while incorporating relevant contextual information. Our Pro3Log approach encodes the history of a process case to a fixed-sized input vector and uses DR-BART, a tree-based ensemble model, for deriving dynamic probability distributions. To handle large event logs, Pro3Log uses a Mixture of Experts framework where several DR-BART models are trained on a predefined number of subsets of an event log. The subsets are obtained by splitting the event log into clusters using the agglomerative Information Bottleneck algorithm, aiming for homogeneity in processing and waiting times within each cluster. This approach reduces data complexity, allowing for more effective training on each cluster. We apply Pro3Log in a business process simulation and demonstrate that it can be used to improve the accuracy of simulation models. These results demonstrate Pro3Log’s potential to enhance diverse BPM techniques, including predictive process monitoring or resource allocation approaches.
