Riess 2022 — Automating Model Management: Metaheuristics for Concept-Drift Adaptation

Open-access survey published in the Journal of Data, Information and Management (Springer), authored during Riess's PhD at NMBU. Forms Paper IV of his PhD thesis.

Summary

The paper reviews literature on using metaheuristics to automate the adaptation of machine-learning models under concept drift. The adaptation problem — retraining, reconfiguring, or reselecting a model when the data-generating distribution shifts — is framed as a high-dimensional optimisation problem that is generally intractable for exact methods; metaheuristics (stochastic, heuristic-based search algorithms not requiring perfect information about problem structure) are therefore the natural fit. Five research questions guide the review: (RQ1) which metaheuristics are used; (RQ2) in what application areas; (RQ3) how metaheuristics assist self-adaptation; (RQ4) which drift types are studied; (RQ5) how evaluation is performed.

Methodologically the review uses Google Scholar retrieval with explicit inclusion/exclusion criteria (including a citation-count threshold for pre-2021 work) and qualitatively codes each study by metaheuristic type, AutoML problem type (feature selection, HPO, FMS), drift type (sudden, gradual, recurring, real vs. synthetic) and evaluation methodology.

Key findings: population-based metaheuristics (GA, PSO, ant colony, bio-inspired variants) overwhelmingly dominate — plausibly because they parallelise naturally and impose few assumptions on problem structure. Usage evolved from single-task automation (feature selection, HPO) toward Full Model Selection (FMS) in online/data-stream settings by 2019–2020. However, evaluation practice is patchy: class distributions are frequently unreported while accuracy is the only metric used, drift types are rarely characterised when real-world data is used, and no study compares multiple population-based metaheuristics against each other on the same drift problem.

Riess recommends that (1) future work treat metaheuristics as models themselves for comparative evaluation, (2) drift characteristics (type, pattern, magnitude) be reported alongside performance, and (3) delayed-label scenarios be evaluated more often since immediate ground truth is uncommon in practice.

The paper is theoretically grounded in the CRISP-DM ML lifecycle (with explicit BPMN modelling of it), stream mining theory (Gama et al. 2013; Bifet & Gavaldà 2007), AutoML (Feurer & Hutter 2019), and ML-lifecycle management (Schelter et al. 2018; Vartak & Madden 2018). Di Francescomarino et al. 2018 (sources/2018-difrancescomarino-genetic-hpo-ppm) and Bose/van der Aalst/Žliobaitė/Pechenizkiy 2011 are cited as BPM-adjacent anchors.

Connections

Feeds concepts/concept-drift as the canonical survey reference in this wiki.
Paper IV of sources/2023-riess-phd-thesis-ppm.
Cited by sources/2025-riess-jorgensen-brage-benchmark-norwegian-llm for the concept-drift motivation of the Norwegian customer-service dialogue benchmark.
Concepts: concepts/concept-drift, concepts/predictive-process-monitoring.
Author: entities/mike-riess.
Related sources: sources/2018-difrancescomarino-genetic-hpo-ppm (GA for PPM HPO).