Open Collections

Abstract

This thesis explores the adaptation of multi-state time-to-event models to regulatory business processes, addressing limitations in traditional process mining techniques when dealing with complex, multi-path processes that exhibit recurrent transitions and time-varying effects. While process mining has evolved significantly with advanced machine learning approaches, many current techniques struggle with competing future pathways, censored observations, and the dynamic nature of real-world processes. We develop a hybrid methodological framework that bridges process mining and time-to-event analysis by combining machine learning for transition path prediction with statistical time-to-event models for duration estimation. This approach effectively models the competing nature of different transitions, accommodates right-censored observations, and handles recurrent events within regulatory workflows. The framework incorporates timespecific Cox models to address time-varying effects of predictors and introduces dynamic quartile selection for personalized predictions tailored to each application’s specific characteristics. Our approach, evaluated on the BC Energy Regulator dataset of 8,461 permit applications through 5-fold cross-validation, shows strong predictive performance with 85.4% accuracy for first transition prediction, 72.6% for full path prediction, and 89.1% for final state prediction. We reduce duration prediction error by 33%, decreasing Mean Absolute Error from 33.1 to 22.3 days. The framework provides interpretable insights through Cox model hazard ratios and Random Forest feature importance metrics, helping stakeholders understand both predictions and their underlying factors. This integration of process mining with time-to-event analysis offers a transparent methodology supporting practical applications in application triage, resource allocation, and process improvement initiatives.