---
title: "Concept Drift"
type: concept
tags: [ml, drift, model-maintenance, automl, data-stream-mining, ppm]
sources:
  - "[[sources/2022-riess-metaheuristics-concept-drift-survey]]"
  - "[[sources/2025-riess-jorgensen-brage-benchmark-norwegian-llm]]"
  - "[[sources/2021-sato-concept-drift-pm-survey]]"
  - "[[sources/2020-baier-handling-concept-drift-bpm]]"
created: 2026-04-20
updated: 2026-05-11
---

# Concept Drift

**Concept drift** (Gama et al. 2013; Žliobaitė et al. 2016) refers to change in distributions and relationships within the data over time. When drift occurs, a deployed machine-learning model can no longer project learned relationships onto the new reality, and predictive performance degrades. Re-training or re-developing models is a costly, often manual operation in production ML systems ([[sources/2022-riess-metaheuristics-concept-drift-survey|Riess 2022]]).

## Taxonomy (Gama et al.)
- **Sudden** — abrupt change in distribution.
- **Gradual** — slow transition.
- **Incremental** — continuous slow drift.
- **Recurring** — seasonality or cyclical return.
- **Real drift** — change in P(y|X).
- **Virtual drift** — change in P(X) without changing P(y|X).
- **Feature drift** — specific feature relevance/distribution changes (Barddal et al. 2017).

## Drift in BPM
Bose, van der Aalst, Žliobaitė & Pechenizkiy (2011) studied drift in [[methods/process-mining-basics|process mining]]. Maisenbacher & Weidlich (2017) explicitly addressed drift in predictive process monitoring. Riess's thesis ([[sources/2023-riess-phd-thesis-ppm]]) flags drift as the *shared vulnerability* of all PPM/PrPM methods: once deployed, any model is subject to drift and needs maintenance.

The PM-specific drift literature is consolidated in [[sources/2021-sato-concept-drift-pm-survey|Sato, de Freitas, Barddal & Scalabrin 2021]] — a 37-page systematic review covering 45 papers across 5 taxonomic axes (type, duration, dynamic, perspective, analysis mode) and 5 challenges (detection, change-point detection, localisation, characterisation, change-process discovery). Sato et al. flag a **reproducibility crisis**: F-score TP/FP/FN definitions vary across studies, no shared benchmark, no agreed detection-delay protocol. They also document that *online* PM remains anaemic (only 7 of 45 papers).

A complementary industrial empirical case is [[sources/2020-baier-handling-concept-drift-bpm|Baier, Reimold & Kühl 2020]] — a P2P process with 70,774 transactions where drift handling (Page-Hinkley + incremental learning + last-batch retraining) lifts throughput-time prediction accuracy from 0.54 to 0.70. Their headline practical recommendation: of three retraining data-selection strategies (`next` / `mixed` / `last`), `last` consistently wins because drift detectors fire with delay and the most recent pre-alarm batch already belongs to the new concept.

See [[syntheses/concept-drift-in-pm]] for the consolidated synthesis combining Sato (survey), Baier (industrial case), and Riess (metaheuristic-adaptation survey).

## Adaptation strategies (Riess 2022 survey)
- **Detection** — Page 1954 CUSUM; Bifet & Gavaldà 2007 EWMA; Pesaranghader & Viktor 2016 Hoeffding drift detection; Ross et al. 2012.
- **Window-based retraining** — retrain on a sliding window of recent data.
- **Ensemble-based** — maintain an ensemble of models, weight/replace by recent performance.
- **Metaheuristic-driven adaptation** — use population-based metaheuristics (Genetic Algorithms, Particle Swarm Optimization, Ant Colony, Harmony Search) to automate feature selection, hyper-parameter optimisation, or full model selection under drift. [[sources/2022-riess-metaheuristics-concept-drift-survey|Riess 2022]] finds these population-based methods dominate the 2012–2020 literature.

## Evaluation pitfalls
Riess 2022 catalogues recurring evaluation issues in the drift-adaptation literature:
- Class distributions often unreported despite accuracy as sole metric (4/17 surveyed studies).
- Drift type and pattern unreported when real-world data is used.
- Ground-truth-immediately-available assumption unrealistic; delayed-label settings underused.
- No head-to-head comparisons of population-based metaheuristics on shared drift tasks.

## Drift in customer service / LLM era
[[sources/2025-riess-jorgensen-brage-benchmark-norwegian-llm|Riess & Jørgensen 2025]] frame drift as the practical motivation for zero-shot LLM classification in customer-service dialogue analysis: the distribution of incoming-call topics changes as products and issues evolve, so retraining supervised classifiers is continuously costly. An LLM invoked via in-context learning with a human-annotator codebook is proposed as a drift-robust (but not yet production-grade) alternative.

## Related
- [[concepts/predictive-process-monitoring]] · [[concepts/process-model-quality]] · [[concepts/mape-k-loop]] (self-adaptation)
- Author: [[entities/mike-riess]]
- See also [[sources/2018-difrancescomarino-genetic-hpo-ppm]] (GA-based HPO in PPM).
- [[syntheses/concept-drift-in-pm]] — synthesis page combining Sato survey + Baier industrial case + Riess metaheuristics.