Concept Drift

Concept drift is the phenomenon in which the relationship a machine learning model learned between its inputs and the target it predicts changes over time, causing the model's accuracy to decline after deployment. A model is trained on historical data under an assumption that future data will resemble the past, but the world changes: customer behaviour shifts, fraud tactics evolve, economic conditions move, and the patterns the model relied on no longer hold. Because a deployed model keeps producing confident predictions even as it becomes wrong, concept drift is a central concern of machine learning operations and model monitoring.

Concept drift versus data drift

Concept drift is often discussed alongside data drift, and the two are related but distinct. Data drift refers to a change in the distribution of the input features the model receives, for example a new age profile of users. Concept drift refers to a change in the mapping from inputs to outputs, meaning the same input now corresponds to a different correct answer. It is possible to have data drift without concept drift, and concept drift without any obvious change in the input distribution. Both can degrade a model, but they call for different monitoring signals and different responses.

| Aspect | Data drift | Concept drift | | --- | --- | --- | | What changes | Input feature distribution | Input-to-output relationship | | Example | New customer demographics | Fraud patterns change meaning of a feature | | Detected via | Distribution comparison | Prediction quality tracking |

Types and causes

Concept drift is commonly categorised by how quickly it unfolds. Sudden drift occurs when the relationship changes abruptly, as when a policy change or external shock instantly alters behaviour. Gradual drift unfolds slowly as an old pattern is progressively replaced by a new one. Incremental drift moves through a continuous sequence of small changes. Recurring or seasonal drift describes patterns that come and go, such as holiday shopping behaviour that returns each year. Identifying the type matters because it influences how often a model should be checked and retrained.

Detection and remediation

Detecting concept drift is easiest when ground-truth labels arrive soon after prediction, because the model's error can then be tracked directly; a rising error rate is strong evidence of drift. When labels are delayed or unavailable, teams rely on proxy signals such as shifts in the distribution of model inputs or outputs, changes in prediction confidence, and statistical tests that compare recent data against a reference window. Dedicated monitoring platforms automate these comparisons and raise alerts.

The primary remedy is retraining the model on fresh, representative data so that it captures the new relationship. Some systems retrain on a fixed schedule, while others retrain only when monitoring detects drift, which is more efficient. Related techniques include online or continual learning, in which the model updates as new data arrives, and shadow deployment, in which a candidate model runs alongside the live model to compare performance before promotion. Careful monitoring is considered essential because an unmonitored model can silently degrade for a long time before the business notices the impact.

Concept drift is especially consequential in Malaysia's regulated financial sector. Credit-scoring and fraud-detection models operated by banks such as Maybank, CIMB, and Public Bank, and by digital banks licensed by Bank Negara Malaysia, must contend with shifting borrower behaviour and evolving fraud schemes. Bank Negara's guidance on the responsible use of artificial intelligence and on model risk management effectively requires institutions to monitor deployed models for degradation, making concept-drift detection a compliance activity as much as an engineering one.

Economic volatility heightens the risk. Events such as the movement-control periods during the pandemic, fuel-subsidy adjustments, and currency fluctuations abruptly changed consumer and business behaviour, producing sudden drift in demand-forecasting, credit, and pricing models across Malaysian retail, telecommunications, and logistics firms including Maxis, TM, and courier operators. Models trained before such shocks required prompt retraining.

Government and public-service AI systems overseen in line with the National AI Office and the Malaysia AI Governance Framework are similarly expected to be monitored for reliability over time, since public-facing predictions that silently degrade can affect service delivery and trust. The Personal Data Protection Act (PDPA) adds a governance layer, since retraining pipelines that reuse personal data must respect consent and purpose limitations.

Building drift-monitoring capability depends on MLOps talent, an area addressed by HRD Corp-supported training and by university data-science programmes at institutions such as Universiti Malaya and Universiti Teknologi Malaysia, as well as the developer ecosystem around MDEC and MRANTI.

1. Gama, J., et al. (2014). A Survey on Concept Drift Adaptation. ACM Computing Surveys.
2. Evidently AI. (2024). What is concept drift in ML, and how to detect and address it. evidentlyai.com.
3. Arize AI. (2024). Model Drift: Concept Drift, Feature Drift, and More. arize.com.