Concept drift
Concept drift is a change over time in the data-generating process, including the distribution of inputs or the relationship between inputs and target outcomes.
Machine Learning System Design
A change over time in the relationship between input features and the target outcome, making previously learned patterns invalid.
- How it is used
- Identifies when a production model’s learned input–outcome relationships no longer support reliable predictions.
- How deep it goes
- Defines the change, distinguishes its patterns, and outlines monitoring and retraining decisions without prescribing a universal remedy.
- Examples
- Uses a UK tax-threshold change, pandemic-driven behavior changes, and changing consumer or economic conditions.
- What you can do after
- Diagnose/debug with it by checking performance and data quality before choosing whether and how to retrain.
Read how Machine Learning System Design explains concept drift →
Financial AI in Practice
A change over time in the relationship between input data and the target a model is intended to predict.
- How it is used
- Frames changing fraud tactics as a reason to monitor and refresh deployed detection models.
- How deep it goes
- Defines drift through evolving fraud and model decay, then gives monitoring and retraining practices without a drift-detection algorithm.
- Examples
- Uses adaptive fraud tactics, new financial products, and newly identified fraud cases to illustrate changing patterns.
- What you can do after
- Operate/configure fraud-model monitoring, investigation feedback, validation, and retraining to respond to changing fraud patterns.
Deep Learning with Python, Third Edition
A change over time in the statistical properties of production data that can degrade a deployed model's performance.
- How it is used
- Frames changing production data as a central reason deployed models lose accuracy and require continuing maintenance.
- How deep it goes
- Explains temporal change, causes, examples, and retraining responses, but does not specify detection metrics or algorithms.
- Examples
- Uses music recommendation, IMDB sentiment classification, and credit card fraud detection to illustrate different drift rates.
- What you can do after
- Operate a maintenance loop that watches production changes, refreshes annotations, and retrains models as their data environment evolves.
Read how Deep Learning with Python, Third Edition explains concept drift →
Designing Deep Learning Systems: A guide for software engineers
A change over time in the relationship between model inputs and the target behavior being predicted.
- How it is used
- Detects production model degradation through continuous monitoring and informs alerts or retraining.
- How deep it goes
- Treats concept drift as a monitoring signal connected to alerting and retraining, without detailing statistical detection methods.
- Examples
- Uses continuous-monitoring pipelines, model-serving APIs, quality gates, developer alerts, and automated retraining as operational examples.
- What you can do after
- Diagnose production degradation with continuous-monitoring pipelines and respond through alerts or model retraining.
Read how Designing Deep Learning Systems: A guide for software engineers explains concept drift →
Deep Learning with R, Third Edition
Concept drift is a change over time in the data-generating patterns or relationships that a deployed machine-learning model depends on.
- How it is used
- Guides dataset collection and maintenance when deployed prediction patterns change over time.
- How deep it goes
- Defines changing production properties, gives domain examples, and stops short of detailed detection methods.
- Examples
- Uses IMDb sentiment across years and rapidly changing credit-card fraud patterns as examples.
- What you can do after
- Diagnose gradual accuracy decay and respond through current-data collection, annotation, and model retraining.
Read how Deep Learning with R, Third Edition explains concept drift →
Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI
Concept drift is a change over time in the relationships between input predictors and the target that can make a previously trained model ineffective.
- How it is used
- Frames model validation and lifecycle planning around changing future data and retraining.
- How deep it goes
- Defines changed predictor-target relationships, explains performance decay, and links drift to retraining without detailing detection.
- Examples
- Uses economic, consumer, social, and political change, plus an XGBoost pipeline retrained through repeated modeling cycles.
- What you can do after
- Operate/configure retraining cycles when predictor-target relationships change and validation results stop representing later performance.
Read how Machine Learning for Tabular Data: XGBoost, Deep Learning, and AI explains concept drift →
Delivering Machine Learning Projects: From design to deployment
Concept drift is a change in the underlying relationships between inputs and outcomes that reduces the applicability of a previously trained model.
- How it is used
- Frames post-deployment model maintenance around changing domains, data flows, interfaces, and infrastructure.
- How deep it goes
- Defines relationship, noise, and interface changes, then recommends monitoring and assertions without detailing adaptation algorithms.
- Examples
- Uses clothing recommendations disrupted by a pandemic and changing data-noise patterns as concrete drift examples.
- What you can do after
- Diagnose model degradation by monitoring performance, checking behavioral assertions, and investigating domain, noise, or interface changes.
Read how Delivering Machine Learning Projects: From design to deployment explains concept drift →
Related concepts
- driftkind of · Drift is a change over time in a system’s behavior, state, data, or configuration relative to a reference or expected baseline.
- stale dataproduced by · Stale data is information that no longer reflects the current state of its underlying source or relevant conditions.
- strategy and monitoring layerused by · The decision-system layer that interprets model signals against business objectives, policies, regulations, and monitored outcomes.
- production datauses · Data generated or collected through real-world operational processes, rather than solely in artificial, simulated, or test settings.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.