A Survey on Concept Drift Adaptation¶
Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A Survey on Concept Drift Adaptation. ACM, 46(4), 1-37.
Cited by¶
22 citations across 22 artifacts.
Each citation links to the sentence it supports in the citing article.
Primes¶
- Calibrated Rule versus Moving World
- The freezing invariant is that \(\hat{f}\) is fixed after training while \(P_{\text{live}}\) drifts, so the gap — formalized as a divergence \(D(P_{\text{live}} \,\|\, P_{\text{train}})\) — grows, and the performance decay is the rise in expected loss tracking that divergence.
This sourceSurveys drift detection and adaptation, factoring the joint distribution and matching update cadence to drift rate.
Supported in partVerified against the work's full text
The survey supports the general idea that a non-adapting model degrades as data change, but does not formalize a divergence D(P_live||P_train) tracking expected loss.
“Predictive models that operate in these settings need to have mechanisms to detect and adapt to evolving data over time, otherwise their accuracy will degrade.”
- The freezing invariant is that \(\hat{f}\) is fixed after training while \(P_{\text{live}}\) drifts, so the gap — formalized as a divergence \(D(P_{\text{live}} \,\|\, P_{\text{train}})\) — grows, and the performance decay is the rise in expected loss tracking that divergence.
- 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.
SupportedVerified against the work's full text
Gama et al. separate virtual drift (p(X) changes, p(y|X) intact), drift in class priors p(y), and real concept drift in p(y|X), which they note is also called concept shift.
“(3) Virtual drift happens if the distribution of the incoming data changes (i.e., p(X) changes) without affecting p(y|X)”
- 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).
- 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.
- It supports inference about refresh cadence: choose retraining frequency by matching the substrate's drift rate rather than by convention or calendar.
- 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.
SupportedVerified against the work's full text
Gama et al. define concept drift as the input-target relation changing over time, and state that predictive models must adapt to evolving data or their accuracy degrades.
“Concept drift primarily refers to an online supervised learning scenario when the relation between the in- put data and the target variable changes over time.”
- This is concept drift
- 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.
Supported in partVerified against the work's full text
The survey establishes concept drift and changes in p(X) as core ML phenomena; it never mentions covariate shift, continual learning or reinforcement-learning reward non-stationarity.
“In dynamically changing and non-stationary environments, the data distribution can change over time yielding the phenomenon of concept drift”
- 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.
- 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.
Supported in partVerified against the work's full text
Supports drift detection, performance metrics such as accuracy, precision and recall, and benchmark comparisons; it does not cover fairness metrics or adversarial testing.
“Predictive models that operate in these settings need to have mechanisms to detect and adapt to evolving data over time, otherwise their accuracy will degrade.”
- 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.
- Reference Cadence Exceeds Tracking Bandwidth
Domain-specific¶
- Covariate-Shift Blind Spot
- The failure is in monitoring posture: operations are configured to watch outcome metrics (recall, precision, AUC) that take weeks or months to materialise as labelled ground truth accumulates
This sourceSurvey of concept drift adaptation, defining drift as change over time in the relation between input data and target variable and cataloguing detection and adaptation strategies.
Supported in partVerified against the work's full text
“Concept drift primarily refers to an online supervised learning scenario when the relation between the input data and the target variable changes over time.”
- The failure is in monitoring posture: operations are configured to watch outcome metrics (recall, precision, AUC) that take weeks or months to materialise as labelled ground truth accumulates
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.
- Its strength is catching distribution shift
- Challenge-Set Refresh Cycle
- The moving target it exists to chase is concept drift — the world changing under a model trained on yesterday's correlations.
This sourceDefines concept drift as an input–target relationship changing over time and surveys adaptation to that moving relation.
- The moving target it exists to chase is concept drift — the world changing under a model trained on yesterday's correlations.
- 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.
- Its strength is catching slow change before it becomes silent failure
- Environmental Shift Retest
- By re-scoring on post-shift evidence and re-checking the proxy, it converts a vague "things feel different now" into a concrete verdict on whether past selections still hold — the textbook response to concept drift, where the distribution a variant was chosen on has moved and its old performance stops predicting.
This sourceThe cited work establishes the textbook response to concept drift, where the distribution a variant was chosen on has moved and its old performance stops predicting.
- By re-scoring on post-shift evidence and re-checking the proxy, it converts a vague "things feel different now" into a concrete verdict on whether past selections still hold — the textbook response to concept drift, where the distribution a variant was chosen on has moved and its old performance stops predicting.
- Model Drift Monitoring
- Its failure modes: label latency means realized-error signals can arrive too late; loose thresholds never trip while tight ones cry wolf; and benign input shifts can masquerade as decay.
This sourceSurveys delayed target feedback, the threshold tradeoff between missed changes and false alarms, and the distinction between real concept drift and input-distribution changes that need not alter the target relation.
- Its failure modes: label latency means realized-error signals can arrive too late; loose thresholds never trip while tight ones cry wolf; and benign input shifts can masquerade as decay.
- Optimization Target Review
- A useful anchor is concept drift — the statistical fact that the relationships a model or metric relies on shift over time.
This sourceDefines concept drift as change over time in the relationships on which a predictive model relies.
- A useful anchor is concept drift — the statistical fact that the relationships a model or metric relies on shift over time.
- Out-of-Distribution Monitor
- Its central blind spot is that watching the inputs catches covariate shift but can miss concept drift
This sourceDefines real concept drift as a change in the input–target relationship that may occur while the input distribution remains unchanged.
- Its central blind spot is that watching the inputs catches covariate shift but can miss concept drift
- Persona Refresh Trigger
- The underlying phenomenon is concept drift
This sourceDefines concept drift as a change over time in the relationship between input data and the target variable.
- The underlying phenomenon is concept drift
- Post-Deployment Validation Monitoring
- The condition it hunts is concept drift — the statistical relationship the pattern relied on changing after deployment, especially where users adapt to the pattern itself.
This sourceDefines concept drift as change over time in the relationship between input data and the target variable.
- The condition it hunts is concept drift — the statistical relationship the pattern relied on changing after deployment, especially where users adapt to the pattern itself.
- Re-grounding Review Cadence
- It is the organizational answer to concept drift, the gradual change in the underlying relationships that quietly invalidates a once-good model.
This sourceDefines concept drift as change over time in the input-target relationship and explains that real drift can be invisible in the input distribution while making the previous model obsolete.
- It is the organizational answer to concept drift, the gradual change in the underlying relationships that quietly invalidates a once-good model.
- Rolling Window Comparison
- Too short a window and normal noise reads as drift, prompting needless churn; too long and real drift is smoothed until it is discovered late — the same concept-drift blind spot the mechanism exists to close, reintroduced through a lazy window choice.
This sourceExplains that concept-drift detectors face a window-size tradeoff between sensitivity to noise and delayed detection of real change.
- Too short a window and normal noise reads as drift, prompting needless churn; too long and real drift is smoothed until it is discovered late — the same concept-drift blind spot the mechanism exists to close, reintroduced through a lazy window choice.
- Selector Retuning Cycle
- The environment shifting out from under a fixed selector is the standard driver, known in machine learning as concept drift, and it is exactly why the loop must keep running rather than converge and stop.
This sourceDefines concept drift as time-varying input–target relations and adaptive learning as updating predictive models online during operation to respond.
- The environment shifting out from under a fixed selector is the standard driver, known in machine learning as concept drift, and it is exactly why the loop must keep running rather than converge and stop.
- Telemetry Drift Dashboard
- Watching against the validated baseline is what turns a raw operations dashboard into evidence about concept drift — the world moving out from under a model that itself never changed.
This sourceDefines concept drift as change over time in the data distribution or predictive relationship while the learned model remains fixed.
- Watching against the validated baseline is what turns a raw operations dashboard into evidence about concept drift — the world moving out from under a model that itself never changed.
- Warrant-Decay Review
- That is warrant decay: the original validation no longer covers the world the model operates in.
This sourceDefines concept drift as change over time in the relation between input data and the target variable, so evidence from the earlier relation may no longer characterize current operation.
- That is warrant decay: the original validation no longer covers the world the model operates in.
Verification¶
Does it exist? Confirmed. This work's DOI resolves to a registered record, which fixes its identity. That is all it fixes.
Does it back the claim? Read against the text for 6 of 22 citations: 2 supported, 4 supported in part. Each verdict is shown under its citation below, with what in the work backs the sentence.
Support is checked per citation rather than per work — the same source can be cited soundly in one article and wrongly in another. Per-citation recording began recently, so a citation with no recorded check is a gap in the record rather than evidence it went unchecked.
See how references were verified.
Registry ID ref:6d5b2c7364d9 · see in the full table