Sparse Mixtures of Shallow Linear Experts for Interpretable and Fast Outcome Prediction

(Process) Outcome Prediction entails predicting a discrete property of an unfinished process instance from its partial trace. Various outcome predictors discovered via Machine Learning (ML) methods, like rule/tree ensembles and (deep) neural networks, have achieved top accuracy performances. However, their opaqueness makes them unsuitable for scenarios necessitating understandable outcome predictors. Aligning with recent efforts to mine inherently interpretable predictors, we suggest training a sparse Mixture-of-Experts, with the ``gate'' and ``expert'' sub-nets being Logistic Regressors. This ensemble of specialized predictors is trained in a end-to-end way while restricting the number of input features used in the sub-nets, as an alternative to typical multi-step/objective mining pipelines (including, e.g., a global feature selection step followed by an ML one). This enables different experts to focus on varied input features for predicting the outcomes of instances in their competency regions. Test results on benchmark logs confirmed the ability of this approach to reach a compelling trade-off between accuracy and interpretability compared to existing solutions.