Concept Drift

An abrupt, discontinuous change in the target relationship, often triggered by an external shock. A new regulation, a competitor's product launch, or a global pandemic instantly invalidates historical patterns.

• Mechanism: A step-function change in the joint distribution P(X,y).
• Detection: Easily caught by control charts or CUSUM algorithms monitoring error rates.
• Example: A fraud detection model becomes ineffective overnight as criminals switch to a completely new attack vector.
• Response: Immediate model rollback or switching to a heuristics-based fallback system.

A gradual, continuous evolution of the target concept over time. The model's accuracy degrades slowly as the underlying process adapts. This is the most common form of drift in production systems.

• Mechanism: A slow, directional change in the decision boundary.
• Detection: Requires sliding window performance metrics or the Drift Detection Method (DDM) to catch before critical failure.
• Example: User preferences for news content slowly shifting over months, making a recommendation engine increasingly stale.
• Response: Scheduled retraining pipelines or online learning with a controlled forgetting factor.

Cyclical or seasonal patterns where the target relationship changes predictably and reverts to a previous state. The concept oscillates between two or more known distributions.

• Mechanism: A periodic shift, often tied to time-of-day, day-of-week, or seasonal buying patterns.
• Detection: Time-series decomposition and periodogram analysis to identify the frequency of recurrence.
• Example: An e-commerce model predicting shopping cart abandonment behaves differently during a Black Friday sale compared to a standard Tuesday.
• Response: Maintaining an ensemble of season-specific models or a context-aware model that conditions on temporal features.

A shift in the input data distribution P(X) that does not necessarily change the conditional target distribution P(y|X). The underlying concept remains valid, but the model encounters unfamiliar data regions.

• Mechanism: Covariate shift without concept shift. The model is extrapolating rather than adapting.
• Detection: Distinguish from real concept drift using two-sample tests on input features, followed by performance analysis.
• Example: A credit scoring model trained on a young demographic suddenly receives applications from a much older population. The scoring logic is still sound, but the input space has expanded.
• Response: Retraining to cover the expanded input domain, rather than changing the core algorithm.

A change in the statistical distribution of the input features (P(X)), not the target relationship. The model sees data unlike its training set.

• Example: A credit model trained on pre-pandemic incomes suddenly receives post-inflation salary data.
• Detection: Population Stability Index (PSI) or Kolmogorov-Smirnov tests on feature distributions.
• Remediation: Retraining on recent data often suffices.

A discrepancy between the data processing logic used during training and the logic used during inference. This is an engineering bug, not a statistical phenomenon.

• Example: A feature is normalized using the full dataset's mean during training, but normalized using a rolling window during serving.
• Detection: Log and compare feature values at both stages.
• Remediation: Use a Feature Store to centralize and version transformation logic.

Performance decay caused simply by the passage of time, often due to concept drift or data drift. It measures the gap between the current model and a hypothetical retrained model.

• Metric: Difference in F1 score or RMSE between the deployed model and a challenger model trained on fresh data.
• Trigger: Automated retraining pipelines should fire when staleness exceeds a business-defined threshold.

A specific type of data drift where the distribution of independent variables changes, but the conditional distribution P(Y|X) remains constant.

• Example: A medical imaging model trained on one hospital's scanner is deployed on a different hospital's newer scanner hardware.
• Remediation: Domain adaptation techniques or importance weighting of training samples.

A change in the distribution of the target variable (Y), independent of the input features. The base rate of what you are predicting changes.

• Example: A fraud detection model faces a sudden spike in fraud rate from 0.1% to 5% during a coordinated attack.
• Detection: Monitor prediction distribution divergence.
• Remediation: Adjust classification thresholds or calibrate probabilities.

A subset of data drift where individual feature statistics change. A feature that was stable becomes volatile, or a feature's range shifts entirely.

• Example: A 'user_age' feature drifts as a platform's demographic ages over a decade.
• Detection: Univariate drift detection per feature using Jensen-Shannon divergence.
• Remediation: Feature selection or engineering to replace unstable features.