Concept
What Is Model Drift?
One-sentence definition
Model drift occurs when a deployed model's real-world performance degrades over time because the relationship between inputs and outcomes has changed since the model was trained.
What real data or case will you use to illustrate it?
The German Credit Lending audit is the anchor. The dataset spans a fixed historical snapshot of German borrowers; the explainer will demonstrate drift by splitting the data temporally — training on the earlier portion, testing on the later portion — and comparing the fairness gap and accuracy metrics across the split using the existing unfair.py pipeline. Employment tenure and age (the proxy and protected attribute identified in the audit) are particularly drift-prone: labour market conditions, job tenure norms, and age-income relationships all shift decade to decade, meaning a model trained in one period may penalise borrowers for features that are no longer predictive — while keeping the demographic disparity intact.
As a second documented case we will use Roberts et al. (2021, Nature Machine Intelligence), which studied hundreds of published COVID-19 diagnostic ML models and found nearly all were unfit for deployment partly due to dataset shift — a peer-reviewed, publicly accessible source.
What are the limitations or trade-offs of this concept?
There are two distinct types — concept drift (the relationship between features and label changes) and data drift (the input distribution shifts) — and they require different responses. Conflating them leads to wrong fixes.
Drift detection requires continuous monitoring infrastructure that most deployments lack; the gap between "drift is happening" and "someone notices" is where harm accumulates undetected.
Retraining on fresh data can correct drift but may introduce new biases if the new data reflects a changed but still discriminatory real-world process — the lending market in a recession encodes different structural disadvantages than the same market in a boom.
Before you start
Concept
What Is Model Drift?
One-sentence definition
Model drift occurs when a deployed model's real-world performance degrades over time because the relationship between inputs and outcomes has changed since the model was trained.
What real data or case will you use to illustrate it?
The German Credit Lending audit is the anchor. The dataset spans a fixed historical snapshot of German borrowers; the explainer will demonstrate drift by splitting the data temporally — training on the earlier portion, testing on the later portion — and comparing the fairness gap and accuracy metrics across the split using the existing unfair.py pipeline. Employment tenure and age (the proxy and protected attribute identified in the audit) are particularly drift-prone: labour market conditions, job tenure norms, and age-income relationships all shift decade to decade, meaning a model trained in one period may penalise borrowers for features that are no longer predictive — while keeping the demographic disparity intact.
As a second documented case we will use Roberts et al. (2021, Nature Machine Intelligence), which studied hundreds of published COVID-19 diagnostic ML models and found nearly all were unfit for deployment partly due to dataset shift — a peer-reviewed, publicly accessible source.
What are the limitations or trade-offs of this concept?
There are two distinct types — concept drift (the relationship between features and label changes) and data drift (the input distribution shifts) — and they require different responses. Conflating them leads to wrong fixes.
Drift detection requires continuous monitoring infrastructure that most deployments lack; the gap between "drift is happening" and "someone notices" is where harm accumulates undetected.
Retraining on fresh data can correct drift but may introduce new biases if the new data reflects a changed but still discriminatory real-world process — the lending market in a recession encodes different structural disadvantages than the same market in a boom.
Before you start