A Leakage-Controlled Prefix-Based Association Rule Mining and Ranking Framework for Predictive Process Monitoring

Predictive process monitoring aims to provide useful insights from ongoing cases to support decision-making. In this context, association rules express interpretable relationships between antecedents and consequents. However, mining retrospective event logs may introduce target leakage, while extensive rule collections can conceal patterns that remain stable over time. This study presents a leakage-controlled pipeline. The proposed pipeline first extracts case prefixes and constructs predictors using only information available within each prefix. It then mines target-constrained rules using FP-Growth on a chronological training partition and ranks the resulting rules using a development-only heuristic score. FP-Growth generated 67 frequent itemset and 19 non-leaking target rules. Of these, 17 rules had a lift higher than one in validation, and 16 of these had a lift higher than one on the untouched test set. Acceptance rules showed the strongest associations. The combination of a medium prefix event volume and a low activity diversity resulted in a test confidence of 0.7047 and a lift of 1.3094 with a support of 0.0289. Cancellation rules showed more moderate associations but remained generally stable; the broadest cancellation rule achieved a test support of 0.2227 and a lift of 1.1028. By contrast, high-rework rules produced only marginal improvements over the 2.19% test prevalence and should therefore be interpreted as weak associations. This study provides a reproducible pipeline for association-rule mining. Its conclusions are limited to observed associations rather than classification accuracy or causal effects