<![CDATA[Non-conformances in organizational processes are often observed as slight irregularities in the sequence of events, prior to even the formal event of a non-conformance being reported, leading to an increased interest in predictive models that can detect such irregularities and give a reason for each alert. This review summarizes the methodological progress of predictive business process monitoring, temporal deep learning and explainable AI and proposes an overall conceptual framework, called temporal explainable AI for compliance-risk prediction, which brings together sequence encoders (e.g. LSTM networks, attention-based transformers) and explainers for post-hoc interpretation (e.g. Shapley additive explanations, local surrogate models, rule-based anchors). The synthesis moves from classical models of stochastic processes to data-driven temporal architectures, laying out the corresponding loss of native interpretability for predictive flexibility, and describing how post-hoc explanation methods have been devised to reclaim the lost interpretability. The results of a comparative review of benchmark evaluations on predictive process monitoring also show that key aspects of the evaluations that are relevant to predictive process monitoring, such as the encoding strategy, cross-log generalization, or handling of class imbalance, have a significant impact on the reported performance and should be taken into account when designing a deployed compliance-risk system. A comprehensive temporal explainable AI system for compliance monitoring should also include concept-drift detection, as the statistical correlation between process behavior and compliance outcomes will not likely be constant throughout an organization's lifecycle.]]>
Copyrights © 2024
