---
title: "Genetic Algorithms for Hyperparameter Optimization in Predictive Business Process Monitoring"
type: source
tags: [ppm, hyperparameter-optimization, genetic-algorithm, outcome-prediction, framework]
authors: [Di Francescomarino Chiara; Dumas Marlon; Federici Marco; Ghidini Chiara; Maggi Fabrizio Maria; Rizzi Williams; Simonetto Luca]
year: 2018
venue: "Information Systems (Elsevier)"
kind: paper
raw_path: "raw/Predictive process monitoring/Genetic algorithms for hyperparameter optimization in predictive business process monitoring.pdf"
created: 2026-04-13
updated: 2026-04-13
key_claims:
  - No single PPM technique under default configuration consistently wins across datasets.
  - A framework with genetic-algorithm and hyperparameter-optimisation layers can automatically select and configure techniques per dataset.
  - Demonstrates scalability and accuracy gains on two real-life logs.
---

# Di Francescomarino et al. 2018 — GA Hyperparameter Optimisation for PPM

Addresses the recurring PPM critique that **no single technique dominates all datasets**. Proposes a framework that chains multiple PPM techniques and **auto-configures** them per-dataset via genetic algorithm HPO.

## Contribution
- Unified framework covering the PPM pipeline (encoding → model → training).
- Two automatic HPO layers — per-step technique selection + per-technique hyperparameter tuning.
- Evaluated on two real-life logs; scalable and competitive with hand-tuned baselines.

## Significance
Part of the **configurability / reproducibility** strand of PPM research. Directly addresses the gripe behind the first sentence of [[sources/2017-tax-lstm-process-prediction|Tax et al. 2017]] ("relative accuracy is highly sensitive to the dataset at hand").

## Cited by
- [[sources/2022-riess-metaheuristics-concept-drift-survey]] — Riess cites Di Francescomarino et al. as the BPM-adjacent anchor for genetic-algorithm HPO and as the closest PPM work addressing earliness (incorporated as a term in the GA fitness function).
- [[sources/2023-riess-temporal-loss-remaining-cycle-time]] — Riess contrasts the GA-based HPO approach (manipulating hyper-parameters to improve earliness) against his own loss-function approach (modifying the training objective directly).

## Connections
**Concepts:** [[concepts/predictive-process-monitoring]] · [[concepts/outcome-prediction]]
**Authors:** [[entities/chiara-di-francescomarino]] · [[entities/marlon-dumas]] · [[entities/fabrizio-maggi]]