Generating multi-level explanations for process outcome predictions

Validation set sampling strategies for predictive process monitoring

“…Gerlach et al introduce an approach that can be used to infer a multi-perspective likelihood graph from (black-box) algorithms trained to predict the next activity [55]. Explanations on multiple levels are provided in recent work by Wickramanayake et al, providing three distinct approaches to determine global explanations of case-level features (with a transparent model), local explanations of event-level features (with an attention-based deep neural network), and local explanations of case-level features (with a novel dual learning deep network) [56]. These explanations might also need evaluation, for which an attempt has been made through well-defined definitions and metrics [36] or through adversarial attacks [57].…”