Concept Drift

What is Concept Drift?

A shift in the underlying statistical properties of data over time, which can lead to a decline in the accuracy of machine learning models.

In other languages

한국어개념 드리프트

시간이 지남에 따라 예측 대상이 되는 데이터의 통계적 특성이 변화하여 AI 모델의 성능이 점차 저하되는 현상입니다.

日本語コンセプトドリフト

時間の経過とともにデータの統計的性質が変化し、AIモデルの予測精度が徐々に低下する現象。

Related Terms

• Data DriftPhenomenon where the statistical properties of input data change over time, causing a model's performance to degrade.
• Agent driftA phenomenon in AI systems where model performance or output accuracy gradually declines over time due to data changes or environmental shifts.
• Model driftThe degradation of a machine learning model's predictive performance over time due to changes in the underlying data distribution or environment.
• Distributional driftA phenomenon where the probability distribution of data encountered by a model during testing or later training stages differs from the distribution it was originally trained on.
• Schema DriftA scenario where the actual structure of data or a tool interface evolves away from the version previously recorded or expected by a system
• Preference driftThe tendency for a user's choices, tastes, or priorities to change over time, requiring an AI to update its understanding.
• Distribution ShiftA phenomenon where the statistical properties of the data used for training differ from the data encountered during deployment or shared between different models.
• Domain shiftA change in the characteristics of input data between the setting where a model was developed and the setting where it is used.
• Covariate ShiftStatistical phenomenon where the distribution of input data changes between the training phase and real-world deployment
• Model collapseA phenomenon where AI models trained on synthetic or AI-generated data suffer from reduced quality and diversity in their outputs over successive generations.
• Capability regressionA phenomenon where an AI model loses its proficiency in previously learned tasks while undergoing training for new information.
• Adversarial inputData specifically crafted to exploit vulnerabilities or trigger incorrect behavior in a machine learning model