---
title: "A Survey on Concept Drift in Process Mining"
type: source
tags: [survey, slr, concept-drift, process-mining, drift-detection, change-point-detection, online-process-discovery, taxonomy]
authors: [Sato, Denise Maria Vecino; de Freitas, Sheila Cristiana; Barddal, Jean Paul; Scalabrin, Edson Emilio]
year: 2021
venue: "ACM Computing Surveys (submission); arXiv:2112.02000. Dec 2021. DOI: 10.1145/3472752"
kind: paper
raw_path: "raw/Process Frameworks & BPM/2021-sato-concept-drift-pm-survey.pdf"
doi: "10.1145/3472752"
arxiv: "2112.02000"
sources: []
key_claims:
  - "Systematic literature review (SLR) of concept drift in process mining: 231 candidates from ACM, IEEE Xplore, Scopus, Springer Link, ICPM Workshops → 45 included after inclusion/exclusion + backward snowballing (88 detection papers + 2 comparison papers, then 3 surveys excluded as ECs)."
  - "Concept drift in PM has TWO branches: (1) drift detection — 38 papers; (2) online PM dealing with evolving environments — 7 papers. PM remains primarily offline."
  - "5-axis taxonomy of drift: TYPE (sudden / gradual / recurring / incremental) · DURATION (momentary / permanent) · DYNAMIC (single-order / multi-order time scales) · PERSPECTIVE (control-flow / time / resource / data / multi-order) · ANALYSIS MODE (offline / online)."
  - "5 challenges in drift detection — drift detection alone · change-point detection · change localisation (which part of the model changed) · change characterisation (what kind of drift) · unravelling change-process evolution."
  - "Evaluation metrics: F-score (geometric mean of precision/recall — but TP/FP/FN definitions vary across 14 of the surveyed papers) and detection delay. NO common evaluation protocol, NO shared benchmark dataset, NO standard metric definition — major reproducibility crisis."
  - "Open challenges: assessment of techniques cumbersome due to lack of shared protocol/data/metrics; online PM techniques rarely consider concept drift; control-flow drift is over-studied vs time/resource/data drift; many approaches detect *only* sudden drift in *only* the control-flow perspective."
created: 2026-05-11
updated: 2026-05-11
---

# Sato, de Freitas, Barddal & Scalabrin 2021 — A Survey on Concept Drift in Process Mining

37-page systematic literature review from the Graduate Program in Informatics (PPGIa), Pontifícia Universidade Católica do Paraná (PUCPR) and IFPR, Brazil. Submitted to ACM Computing Surveys (December 2021). The wiki's anchor for **concept drift specifically within process mining** — distinct from but complementary to [[sources/2022-riess-metaheuristics-concept-drift-survey|Riess 2022]] (which surveys metaheuristics for drift *adaptation* across ML broadly).

## What the survey does

PRISMA-style SLR over four digital libraries + ICPM workshops:

| Source | Hits |
|---|---|
| ACM Digital Library | 20 |
| IEEE Xplore | 13 |
| Scopus | 56 |
| Springer Link | 140 |
| ICPM Workshops | 2 |
| **Total** | **231** |

After inclusion/exclusion criteria (IC1 peer-reviewed empirical drift-PM papers + IC2 comparison papers; ECs filtering duplicates, papers with <5 citations pre-2015, perspective-incomplete papers, weakly-documented experimental papers) and backward snowballing of 3 surveys → 45 papers analysed.

Three research questions: (RQ1) what drift-detection approaches; (RQ2) what tools; (RQ3) how validated.

## The taxonomy (Figure 3 in the paper)

Concept drift in PM organised along five axes:

| Axis | Categories |
|---|---|
| **Type** | Sudden · Gradual · Recurring · Incremental |
| **Duration** | Momentary · Permanent |
| **Dynamic** | Single-order · Multi-order (changes at different time granularities) |
| **Perspective** | Control-flow · Time · Resource · Data · Multi-order |
| **Analysis mode** | Offline · Online |

13 structural change patterns are referenced (Weber et al. 2008; citation [77] in the survey) for imperative process models: adding/deleting fragments, moving/replacing fragments, adding/removing sub-process levels, adapting control dependencies, changing transition conditions.

## Five challenges in drift handling

1. **Drift detection** — has a process changed?
2. **Change-point detection** — when did it change?
3. **Change localisation** — which region of the model changed?
4. **Change characterisation** — what type of drift, in which perspective?
5. **Unravelling change-process evolution** — discovering the *drift process* itself (meta-discovery).

## Branches and breakdown

| Approach | Papers |
|---|---|
| Concept drift detection | 38 |
| Online PM dealing with evolving environments | 7 |

## Open problems flagged

- **Evaluation crisis**: F-score is the dominant metric but TP/FP/FN definitions vary across 14 of the surveyed papers; no shared benchmark dataset; no agreed detection-delay protocol → comparing techniques is "cumbersome."
- **Perspective imbalance**: control-flow drift dominates; time, resource, and data drift are under-studied.
- **Online PM is anaemic**: only 7 papers; online PM techniques rarely treat drift as a first-class concern.
- **Drift-type coverage**: many methods detect only *sudden* drift; gradual, recurring, incremental drift remain hard.

## Why this matters for the wiki

- Provides the canonical taxonomy + open-problem inventory for [[concepts/concept-drift]] in PM contexts.
- Complementary to [[sources/2022-riess-metaheuristics-concept-drift-survey]] — Sato surveys *detection*; Riess surveys *adaptation*.
- Complementary to [[sources/2020-baier-handling-concept-drift-bpm]] — Sato is the meta-level survey; Baier is the empirical industrial case.
- Sets baseline expectations against which any future drift-aware PPM/LLM-PPM work should be evaluated.

## Limitations

- 2021 cut-off — post-2021 work on Transformer-based drift detection, LLM-as-drift-detector, online stream-PM with foundation models is not covered.
- Excludes some industrial / non-peer-reviewed work.
- The SLR's own exclusion criterion EC3 (approaches not considering changes in a complete perspective of the model) excludes some PPM-specific drift work where the target is a process *property* rather than a model — relevant given that PPM is often property-targeted.

## Connections

**Concepts:**
- [[concepts/concept-drift]] — *anchor reference*.
- [[concepts/predictive-process-monitoring]] — drift handling is a deployment requirement for PPM.

**Methods:**
- [[methods/process-discovery-methods]] — drift detection sits adjacent to discovery; the 7 "online PM" papers are discovery papers handling streams.
- [[methods/process-mining-basics]] — survey scope is broader than discovery (also conformance + enhancement).

**Sources:**
- [[sources/2022-riess-metaheuristics-concept-drift-survey]] — sibling survey on adaptation.
- [[sources/2020-baier-handling-concept-drift-bpm]] — empirical industrial case study.
- [13] (Bose, van der Aalst, Žliobaitė, Pechenizkiy 2011 / 2013) "Dealing with concept drifts in process mining" — *priority referenced source; not in raw/* but the foundational drift-in-PM paper.
- [16] (Bifet & Gavaldà 2007) ADWIN — referenced ML-side foundation.

**Syntheses:**
- [[syntheses/concept-drift-in-pm]] — *anchor source*; pairs with Riess metaheuristics + Baier industrial case.
- [[syntheses/ppm-landscape]] — augments the open-problem list in §6.