« Back to Glossary Index

Concept drift is led in dynamic and non-stationary environments, where data distributions can change over time. Concept drift refers to changes in the conditional distributions of what is typically defined as the target variable, given a set of input values the distributions of which remain unchanged. For example, in a classification task, concept drift can be formally described as the disparity in joint distributions between the target (y) and features (X) measured between time t0 and time t1, where the prior probabilities of classes or the class conditional probabilities between X and y may change.

A typical example of the concept drift is when a news reader changes interest when following an online news stream. Whilst the distribution of the incoming news are the same, the conditional distribution of the interesting news documents for that reader may change.