Generative Artificial Intelligence (GenAI) techniques have recently emerged as promising tools for challenges in Process Mining such as missing data, small datasets, and the generation of complex what-if scenarios. A major limitation of current approaches is their inability to produce sets of process traces in a mutually dependent manner. Traces are generated independently, failing to reflect crucial process-level phenomena such as concurrency, resource contention, and workload distribution. This limitation significantly reduces the realism and utility of generated logs, yet has received limited attention in the literature. In contrast, Business Process Simulation (BPS) excels at preserving process constraints and interdependencies. However, in the presence of specific or complex what-if scenarios, this requires the adjustment of several simulation parameters, which can be challenging to achieve. Motivated by these issues, where generative methods lack consistency and BPS methods lack flexibility, we propose a novel data generation framework combining GenAI with BPS. The proposed approach enables generating diverse and complex whatif scenarios via data-driven generative models, which are then validated and refined through BPS to ensure consistency and feasibility at the log level. We call the approach What-If Scenarios via Data-driven generative mOdels and siMulation (WISDOM). We compare several existing methods for process data generation and assess their outputs using an established evaluation framework. Results show that while generative methods fail to capture the full intricacies of process behaviour, their integration with BPS produces traces that are both diverse and compliant with process constraints. This hybrid approach addresses a critical gap in the literature and opens new avenues for robust, simulation-grounded generation methods in process mining.

Recent research in Computer Science has investigated the use of Deep Learning (DL) techniques to complement outcomes or decisions within a Discrete Event Simulation (DES) model. The main idea of this combination is to maintain a white box simulation model complement it with information provided by DL models to overcome the unrealistic or oversimplified assumptions of traditional DESs. State-of-the-art techniques in BPM combine Deep Learning and Discrete Event Simulation in a post-integration fashion: first an entire simulation is performed, and then a DL model is used to add waiting times and processing times to the events produced by the simulation model. In this paper, we aim at taking a step further by introducing RIMS (Runtime Integration of Machine Learning and Simulation). Instead of complementing the outcome of a complete simulation with the results of predictions a posteriori, RIMS provides a tight integration of the predictions of the DL model at runtime during the simulation. This runtime-integration enables us to fully exploit the specific predictions while respecting simulation execution, thus enhancing the performance of the overall system both w.r.t. the single techniques (Business Process Simulation and DL) separately and the post-integration approach. In particular, the runtime integration ensures the accuracy of intercase features for time prediction, such as the number of ongoing traces at a given time, by calculating them during directly the simulation, where all traces are executed in parallel. Additionally, it allows for the incorporation of online queue information in the DL model and enables the integration of other predictive models into the simulator to enhance decision point management within the process model. These enhancements improve the performance of RIMS in accurately simulating the real process in terms of control flow, as well as in terms of time and congestion dimensions. Especially in process scenarios with significant congestion – when a limited availability of resources leads to significant event queues for their allocation – the ability of RIMS to use queue features to predict waiting times allows it to surpass the state-of-the-art. We evaluated our approach with real-world and synthetic event logs, using various metrics to assess the simulation model's quality in terms of control-flow, time, and congestion dimensions.

In recent years, trace generation has emerged as a significant challenge within the Process Mining community. Deep Learning (DL) models have demonstrated accuracy in reproducing the features of the selected processes. However, current DL generative models are limited in their ability to adapt the learned distributions to generate data samples based on specific conditions or attributes. This limitation is particularly significant because the ability to control the type of generated data can be beneficial in various contexts, enabling a focus on specific behaviours, exploration of infrequent patterns, or simulation of alternative "what-if" scenarios.In this work, we address this challenge by introducing a conditional model for process data generation based on a conditional variational autoencoder (CVAE). Conditional models offer control over the generation process by tuning input conditional variables, enabling more targeted and controlled data generation. Unlike other domains, CVAE for process mining faces specific challenges due to the multiperspective nature of the data and the need to adhere to control-flow rules while ensuring data variability. Specifically, we focus on generating process executions conditioned on control flow and temporal features of the trace, allowing us to produce traces for specific, identified sub-processes. The generated traces are then evaluated using common metrics for generative model assessment, along with additional metrics to evaluate the quality of the conditional generation.