Causal Process Mining: From Interpretability to Causal Explainability in Temporal Deviations

In operations with timing constraints, drifts in execution times raise costs, weaken SLAs, and degrade user experience. While process mining reveals where and when deviations appear, correlation-based analyses rarely explain why they happen or what to change. We present a causal process mining method that moves from interpretability to causal explainability. First, we localize temporal deviations per process variant using performance spectra and cluster traces with similar interval profiles; then we learn a consensus causal graph (LiNGAM, PC, and GES) constrained by domain knowledge to make relationships readable and auditable. Second, on that graph we run intervention queries (do-operator) to answer “what-if” questions with identification conditions and uncertainty summaries, turning signals into actionable recommendations. We validate the approach on an event log derived from planning artifacts (1,261 traces; 6,015 events; 8 activities; 6 variants). The analysis uncovers causal routes that link complexity level and number of requirements to development/testing hours and meetings, and quantifies the expected impact of alternative policies on those outcomes. The method contributes: (i) variant-level causal graphs that preserve process heterogeneity; (ii) robust effect estimation that handles hidden heterogeneity across variants; (iii) spectrum-guided temporal localization of interventions; and (iv) explicit, auditable recommendations with confidence information to reduce temporal deviations.