Overview
Direct Answer
Data drift refers to the degradation of machine learning model performance caused by shifts in the statistical distribution of input features or target variables after deployment. This phenomenon occurs when the real-world data generating process diverges from the training data, violating the assumption that training and production distributions remain constant.
How It Works
Models learn patterns from historical training data and optimise weights based on those distributions. When production data exhibits different feature correlations, class proportions, or value ranges, the model's learned decision boundaries become misaligned with actual patterns. This misalignment accumulates as predictions become increasingly inaccurate without explicit retraining or monitoring mechanisms to detect distributional changes.
Why It Matters
Model degradation directly impacts business outcomes through reduced prediction accuracy, flawed decision-making, and compliance violations in regulated industries. Organisations that fail to detect and remediate drift experience financial losses, customer dissatisfaction, and reputational damage. Continuous monitoring and retraining are essential to maintain model reliability and ROI.
Common Applications
Fraud detection systems experience drift as fraudster behaviour evolves; credit risk models drift when economic conditions shift; recommendation engines drift as user preferences change seasonally; medical diagnostic models drift as patient demographics or equipment calibration varies.
Key Considerations
Distinguishing data drift from concept drift (target distribution changes) requires different remediation strategies. Drift detection introduces operational overhead and latency considerations that must be balanced against the cost of model degradation.
Cross-References(1)
Referenced By1 term mentions Data Drift
Other entries in the wiki whose definition references Data Drift — useful for understanding how this concept connects across Data Science & Analytics and adjacent domains.
More in Data Science & Analytics
Real-Time Analytics
Applied AnalyticsThe discipline of analysing data as soon as it becomes available to support immediate decision-making.
Customer Analytics
Applied AnalyticsThe practice of collecting and analysing customer data to understand behaviour, preferences, and lifetime value.
Geospatial Analytics
VisualisationThe analysis of geographic and spatial data to discover patterns, relationships, and trends tied to location.
Diagnostic Analytics
Statistics & MethodsAnalysis techniques focused on understanding why something happened by examining data patterns and correlations.
MLOps
Statistics & MethodsThe practice of collaboration between data science and operations to automate and manage the machine learning lifecycle.
Data Wrangling
Statistics & MethodsThe process of cleaning, structuring, and enriching raw data into a desired format for analysis.
Big Data
Statistics & MethodsExtremely large and complex datasets that require advanced computational tools and techniques to store, process, and analyse.
Churn Analysis
Applied AnalyticsThe process of analysing customer attrition to understand why customers stop using a product or service.