Process mining extracts models from event logs to provide insights into workflows. Current techniques often struggle to capture counting-based dependencies between activities. We address this challenge by introducing guarded process trees, an enhancement of traditional process trees with guards that control transitions based on activity occurrence counts. We present a two-step method for translating guarded process trees to Petri nets: first, converting to a guarded Petri net, then removing the guards while preserving the counting semantics. We formally define and prove the correctness of both translation steps. To demonstrate practical applicability, we develop a method for learning guarded process trees from event logs and implement it using grammatical evolution. Through experiments on both synthetic and real-world datasets, we show that our approach significantly improves precision compared to existing techniques while maintaining model interpretability. The evaluation demonstrates average precision improvements of $15.31 \%$ across datasets when enhancing the Inductive Miner with guards, with individual improvements ranging from $2.06 \%$ to $70.51 \%$.
Predictive Process Monitoring (PPM) aims to forecast the future behavior of an ongoing process instance based on historical event logs. PPM approaches which follow an offline training and online prediction paradigm, often struggle to adapt to the evolving nature of real-world processes. While state-of-theart approaches address such challenges by continuously updating or retraining models, they typically overlook dynamic processes, where numerous unseen activities emerge or the order between activities evolves over time. To address this gap, we propose a Class Incremental Learning with Drift Detection and Data Augmentation (CIL2D) framework for next activity prediction in dynamic process environments. CIL2D employs a representationbased drift detection mechanism to identify shifted, novel, or unseen activities. To enhance model generalization to drifting patterns, CIL2D applies data augmentation in feature space to generate diverse and novel feature representations that are unseen but likely to occur as processes evolve. Upon drift detection, the model is incrementally updated to the newly observed data in combination with augmented samples and replay buffer traces. Experiments on real-life event logs demonstrate that CIL2D consistently outperforms existing methods in prediction accuracy and adaptation efficiency, highlighting its effectiveness for next activity prediction in highly dynamic process environments.
