Metzger et al. 2015 — Comparing and Combining PPM Techniques

Early empirical survey / ensemble paper. Compares ML-based (classifier on running features), constraint-based (QoS aggregation over remaining sub-processes), and combined techniques on a real transport-logistics log.

Contribution

Categorisation of prior PPM families into ML vs constraint-satisfaction vs combined.
Evaluation on 3942 process instances, five months of real data.
Demonstrates ensembling / combining techniques outperforms either alone — an early instance of the "no single best model" thesis that Tax et al. 2017 later generalised with a unified LSTM.

Significance

Represents the pre-deep-learning PPM era that neural methods would largely supersede from 2017 onward. Remains relevant where interpretability, small-data regimes, or constraint guarantees are required. Andreas Metzger (entities/andreas-metzger) continues to publish on related self-adaptive process monitoring — cited extensively in the APM Manifesto.

Connections

Concepts: concepts/predictive-process-monitoring · concepts/outcome-prediction Authors: entities/andreas-metzger