Concept Drift

Definition

Concept Drift is a machine learning term for a change over time in the relationship between input data and the outcome a model predicts, so that a model trained on past data becomes less accurate. It is related to data drift, where the distribution of inputs changes.

Why it matters

Models learn patterns from past data. When user behaviour, language, markets or meanings change, models need to be monitored and updated to remain accurate.

How it works

Types of drift

  • Drift can be sudden, gradual, incremental or recurring, such as seasonal changes.

Detection

  • Drift is detected by monitoring model performance and changes in data distributions over time.

Response

  • Models are retrained, updated or supplemented with current data, such as through retrieval.

How Netsleek uses the term

Netsleek uses Concept Drift as an analogy when meanings in a market change, for example when a category term shifts in meaning. It monitors whether AI systems describe client categories in outdated ways.

Comparisons

  • Concept Drift vs Semantic Stability: Semantic stability is meaning remaining consistent. Concept drift is meaning or relationships changing over time.
  • Concept Drift vs Signal Decay: Signal decay is particular signals weakening. Concept drift is a change in underlying relationships.

How Concept Drift Connects Across AI Systems

Concept Drift does not operate in isolation. It connects to other concepts across the AI search and visibility process:

Related glossary concepts

Common misinterpretations

  • Concept drift is not the same as a model bug.
  • It does not affect all tasks equally.
  • Netsleek’s application to market meanings is an analogy, not the technical definition.

Summary

Concept Drift is the change over time in the relationships a model has learned. It requires monitoring and updating to keep models accurate.