Concept drift occurs when the patterns a machine learning model learned during training no longer reflect reality, causing its predictions to become less accurate over time.
A spam detection model trained on older email patterns, for example, may struggle as spammers evolve their techniques to mimic legitimate messages.
Concept drift comes in several forms.
- Gradual: Patterns shift slowly over time, causing a steady decline in model performance.
- Sudden: An abrupt change in the environment, such as a major market disruption or a global event like a pandemic, makes the model's learned patterns obsolete almost immediately.
- Recurring: Changes happen in a predictable cycle, such as higher sales during holidays or different user behavior on weekends.
Concept drift can be detected by monitoring a model’s performance metrics over time. A drop in accuracy, precision, or recall can signal that drift is occurring.
The most common response to concept drift is retraining the model on more recent data so it can relearn the current patterns.
Share