A Survey on Concept Drift Adaptation.

Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A Survey on Concept Drift Adaptation. ACM Computing Surveys, 46(4), 1-37.

Type

Journal article

Intellectual base

Review or monograph

Year

2014

DOI

10.1145/2523813

Link

https://doi.org/10.1145/2523813

Cited from

computer_science, data_science, engineering_design, philosophy, systems_cybernetics

Cited by

9 citations across 9 artifacts.

Each citation links to the sentence it supports in the citing article.

Primes

• Calibrated Rule versus Moving World
  • The same structure governs a fraud-detection model decaying as fraudsters adapt (concept drift), a demand-forecasting model failing when a new product category shifts the input mix (data drift), and a churn model going stale after a pricing change rewrites customer behavior (regime break).

    This sourceSurveys drift detection and adaptation, factoring the joint distribution and matching update cadence to drift rate.

• Concept Drift
  • The input distribution can shift while the input-to-outcome mapping holds (covariate shift); the base rate of outcomes can shift (prior shift); the very relationship between inputs and outcomes can shift (concept shift, the genuine article).

    This sourceComprehensive survey distinguishing drift modes (real concept drift vs. virtual / covariate / prior shift) and cataloguing detection and adaptation countermeasures keyed to those modes.

• Data Drift
  • It supports inference about refresh cadence: choose retraining frequency by matching the substrate's drift rate rather than by convention or calendar.

    This sourceSurveys drift detection and adaptation, including matching refresh/retraining cadence to the rate at which the substrate moves.

• Foreseeing (Prediction)
  • This is concept drift

    This sourceCanonical survey of concept drift — when the relation between inputs and target changes over time — and of adaptive/continual-learning strategies (changepoint detection, ensembles upweighting recent data); supports the T6 claim that models trained on historical data degrade under concept drift and the ML terminology used.

• Non-Stationary Objective
  • In machine learning, concept drift, covariate shift, and distribution shift underlie the entire continual-learning subfield, and reward-function non-stationarity is a recognized failure mode in reinforcement learning.

    This sourceSurveys concept drift, covariate shift, and distribution shift underlying continual learning; the machine-learning instance of a non-stationary objective.

• Quality Control
  • AI and machine learning: Model evaluation against performance metrics (accuracy, precision, recall, fairness), adversarial testing, benchmark comparisons, drift detection (model behavior degradation over time), as Gama et al. (2014) survey in their canonical treatment of concept drift adaptation in evolving data streams.

    This sourceCanonical survey of model-monitoring and drift-detection techniques for evolving data streams; the reference treatment for AI/ML quality control over time.

• Reference Cadence Exceeds Tracking Bandwidth
  • … though which lever is cheapest in which medium is this entry's own ranking rather than a measured comparison: in software, working in small batches creates shorter lead times and faster feedback loops; in ML, a model not adapted as the distribution moves degrades in accuracy, so retraining cadence is the lever;

Mechanisms

• Adaptive Policy Review Cycle
  • Its strength is catching distribution shift

    This sourceSurveys adaptive-learning strategies that revise models in response to statistically detected concept drift.

• Drift Sample Review
  • Its strength is catching slow change before it becomes silent failure

    This sourceExplains that gradual or incremental concept drift can make an existing classifier obsolete and erode accuracy unless the change is detected promptly.

Verification

This reference passed the adversarial substantiation pipeline: it was checked to exist and to support the claim it is attached to. See how references were verified.

Registry ID ref:6d5b2c7364d9 · see in the full table