Abstract
Predictive Process Monitoring (PPM) deals with providing predictions about the continuation of partially executed process executions based on historical process data. PPM techniques have been developed using increasingly complex Machine and Deep Learning architectures, which lack interpretability of the predictions. Recently, explainable PPM techniques have been proposed, thus making them more ”trustable” for the users. Amongst these techniques, counterfactuals aim at suggesting, for a given process execution, the minimal changes to be applied to it to achieve a desired outcome. In this paper, we introduce an evaluation framework for evaluating different approaches for the generation of counterfactuals in PPM. The framework is used to evaluate these approaches against several real-life datasets. The results show that, although a clear winner cannot be identified, each approach is suitable for logs with specific characteristics, or for achieving specific objectives.