DIFF-ERO: A Conformance-Aware Loss for Deep Learning in Process Mining

DIFF-ERO is a newly introduced conformance-aware loss function designed for deep learning models applied to process data. It leverages a differentiable formulation of stochastic conformance that integrates control-flow information, enabling batch-level stochastic transition matrices to enhance structural precision and recall during training. This model-agnostic approach, instantiated in transformer encoder-decoder architectures, demonstrates improved predictive performance in scenarios where structural accuracy is critical, while also aligning the learned stochastic automaton with the underlying process model structure, making it significant for practitioners focused on process mining and predictive analytics.