Synthesis: Concept Drift in Process Mining and PPM

How three angle-different surveys + one industrial case + one emerging-method paper together describe the state of concept drift in BPM as of May 2026.

The three surveys are non-overlapping

Survey	Scope	What it surveys	Year
Sato et al. 2021	Concept drift in process mining	Detection + online PM under drift — 45 papers, 5-axis taxonomy (type, duration, dynamic, perspective, analysis mode), 5 challenges (detection, change-point detection, localisation, characterisation, change-process discovery)	2021
Riess 2022	Metaheuristics for drift adaptation (cross-domain, not BPM-specific)	Population-based metaheuristics (GA, PSO, etc.) for retraining/reselecting models under drift across engineering, finance, social science	2022
Baier, Reimold & Kühl 2020	Empirical industrial case study	P2P (procure-to-pay) process at a large German company; 70 k transactions 2016–2018; Naïve Bayes + Page-Hinkley/ADWIN + 3 retraining strategies	2020

The triangulation: Sato tells us what to detect, Riess tells us how to adapt, Baier tells us what works in production.

Convergent findings

Drift handling works — if you do it

Baier's headline result: static-baseline 0.5400 → drift-handled 0.7010 accuracy on throughput-time prediction. This is a +30 percentage-point absolute improvement from continuous monitoring + adaptation, on a real industrial event log.

`last` strategy outperforms `next` and `mixed`

Baier compares three data-selection strategies for retraining the model after a detected drift:

next — retrain on the batch after drift detection
mixed — retrain on data spanning the detection
last — retrain on the last batch before drift detection

last wins consistently across both detectors (Page-Hinkley, ADWIN) and all four batch sizes (500, 1000, 2000, 5000). Intuition: drift detectors fire with lag, so the most recent pre-alarm batch already belongs to the new concept.

Small batches outperform large

Baier finds batch size 500 > 1000 > 2000 > 5000 across detectors. Larger batches are too slow to adapt to subsequent drift. Suggests a deployment heuristic: pick the smallest batch size your inference latency budget allows.

Drift detector choice matters less than retraining-strategy choice

Page-Hinkley narrowly beats ADWIN in Baier, but the strategy gap is wider than the detector gap. The data-selection question dominates the detection-algorithm question for practical performance.

Divergent findings / open tensions

Sato's reproducibility crisis vs. Riess's evaluation rigour critique

Both flag evaluation gaps but from different angles:

Sato: F-score TP/FP/FN definitions vary across 14 of the surveyed drift-detection papers; no shared benchmark; no agreed detection-delay protocol.
Riess: 4 of 17 metaheuristic-adaptation studies don't report class distribution despite using accuracy as sole metric; drift type/pattern unreported when real-world data used; no head-to-head comparison of population-based metaheuristics on the same drift task.

Synthesis: the drift-in-PM literature suffers from two parallel reproducibility crises — one on the detection side, one on the adaptation side. Both surveys converge on the recommendation: publish drift characteristics + class distributions + comparable evaluation protocols.

Online vs. offline

Sato documents that only 7 of 45 surveyed PM papers handle drift online; the remaining 38 are offline detection. Baier's industrial case operates near-online (stream-of-batches, not stream-of-events). PPM deployments overwhelmingly need online or near-online behaviour, yet the literature is dominated by offline analysis.

Control-flow drift vs. multi-perspective drift

Sato's perspective taxonomy distinguishes control-flow, time, resource, data, and multi-order drift. The literature is heavily biased toward control-flow drift; time/resource/data drift (which Baier's case exemplifies — an automation-level feature shift) is under-served.

Implications for LLM-PPM

Padella et al. 2026 do not directly test drift robustness, but several of their findings are drift-relevant:

Open question: how do LLM-PPM predictions degrade under drift, particularly when the drift contradicts strong pre-training priors? Empirical work missing.

Reading order

Sato et al. 2021 — taxonomy & open-problem inventory.
Baier, Reimold & Kühl 2020 — concrete industrial case + retraining-strategy lessons.
Riess 2022 — adaptation toolkit (metaheuristics).
Padella et al. 2026 — drift-adjacent implications of LLM-PPM.

Open threads

Maaradji et al. 2017 (sudden/gradual drift detection from execution traces) — referenced in Sato + Baier, not yet in raw/.
Bose, van der Aalst, Žliobaitė & Pechenizkiy 2011 (foundational PM-drift paper) — referenced everywhere, not yet in raw/.
Ostovar et al. 2020 (robust drift characterisation) — referenced in Sato, not in raw/.
Yeshchenko et al. 2019 (visual-analytics drift detection) — referenced in Sato, not in raw/.
LLM-PPM + drift: no ingested paper combines the two empirically.
Causal drift — when is a detected drift intervention-worthy vs. spurious? Currently treated as detection-only.

Concept hub: concepts/concept-drift.
PPM hub: concepts/predictive-process-monitoring.
LLM-PPM hub: concepts/llm-based-ppm.
Landscape: syntheses/ppm-landscape.
Reading list: syntheses/llm-bpm-reading-list.