A Reproducible Protocol for Resource-Aware Predictive Process Monitoring: Compact Baselines, a Simulator Blueprint, and Pitfalls

Agents4Science 2025 Conference Submission341 Authors

Keywords: Predictive Process Monitoring, Resource-Aware Baselines, AI Scientist

TL;DR: A reproducible protocol for predictive process monitoring with compact LSTM baselines and a lightweight resource-aware simulator blueprint, highlighting both benefits and pitfalls of resource-centric approaches.

Abstract: We present resource-aware predictive process monitoring (PPM) as a modular, agent-based design that complements case-centric next-activity predictors with explicit modeling of shared resource contention. Our contributions are fourfold. (i) A leakage-safe, deterministic protocol with chronological case splits, train-only normalization, fixed seeds, and automatic artifact logging. (ii) A compact, transparent LSTM baseline for next-activity prediction on three public logs (BPI 2012, BPI 2017, Road Traffic) with ready-to-reuse splits and scripts. (iii) A released simulator blueprint with per-resource multinomial policies and lightweight discrete-event simulation, plus evaluation measures spanning global next event, workload MAPE, and per-resource next-task precision. (iv) Pitfalls and checklists observed in practice (e.g., lifecycle pairing under partial traces; imbalance-aware back-offs). Baseline next-activity results are strong (Top-3 0.987--0.994; Top-1 0.757--0.833), exposing systematic confusions that motivate resource context. Code, splits, and plot artifacts enable one-click replication. This paper is intended as a protocol + baseline + blueprint to accelerate trustworthy resource-aware PPM experiments; we do not claim state-of-the-art accuracy nor report end-to-end simulator metrics in this version.

Submission Number: 341