New Loss Function Improves Deep Learning for Process Mining Conformance.

Deep learning has significantly advanced process analytics, particularly in predictive and prescriptive monitoring. However, conventional objectives like cross-entropy primarily optimize local next-step likelihoods, often failing to adequately capture the global control-flow structure of processes. This can lead to models with high token-level accuracy but imprecise overall behavior. To address this, a new method called DIFF-ERO is proposed. It is a conformance-aware loss function designed for deep learning models that process event data. DIFF-ERO offers a differentiable approach to entropy-based stochastic conformance, directly integrating control-flow information into the training process. The method constructs batch-level stochastic transition matrices with soft edge memberships, allowing structural precision and recall signals to guide backpropagation. DIFF-ERO is model-agnostic and can be applied to any model parametrizing stochastic transitions. When instantiated in transformer encoder-decoder pipelines for next-activity prediction, it shows improved predictive performance where process structure is critical, while also ensuring the learned stochastic automaton converges towards the true process model structure.