AI 용어집
인공지능 완전 사전
Concept Drift
Change in the conditional probability distribution P(y|x) where the relationships between features and the target evolve over time, requiring adaptation of predictive models.
Virtual Drift
Change in the marginal distribution P(x) of input features without modification of the underlying relationship between features and the target P(y|x).
Real Drift
Fundamental modification of the relationship between features and the target variable P(y|x), directly affecting the model's prediction performance.
DDM (Drift Detection Method)
Statistical monitoring algorithm based on model error rates using the binomial distribution to detect significant changes in performance.
EDDM (Early Drift Detection Method)
DDM variant optimized to detect gradual changes by monitoring the average distance between successive errors rather than just the error rate.
ADWIN (Adaptive Windowing)
Adaptive algorithm that maintains a sliding window of variable size and statistically compares the distributions of two sub-windows to detect changes.
Page-Hinkley Test
Statistical change detection test based on the accumulation of differences between observed values and their mean, effective for identifying abrupt drifts.
KS Test (Kolmogorov-Smirnov Test)
Non-parametric test comparing the cumulative distribution functions of two samples to determine if they come from the same distribution.
EWMA (Exponentially Weighted Moving Average)
Statistical smoothing method that assigns exponentially decreasing weights to older observations to detect changes in time series.
Drift Detection Rate
Metric measuring the proportion of actual drifts correctly identified by a detection algorithm, thus evaluating its detection effectiveness.
False Alarm Rate
Proportion of false drift detections reported by an algorithm when no actual change has occurred in the data distribution.
Detection Delay
Time elapsed between the actual occurrence of a concept drift and its detection by the monitoring algorithm, measuring the system's responsiveness.
Window-based Methods
Drift detection approaches using temporal windows to compare statistical distributions between recent and historical data.
Statistical Process Control
Set of statistical methods used to monitor and control processes, adapted for detecting drifts in data streams.
Change Point Detection
Identification of precise moments when the statistical properties of a time series change significantly, fundamental for drift detection.
Feature Drift
Change in the distribution of one or more input features P(xi) over time, which can indirectly affect model performance.
Prior Probability Drift
Change in the marginal distribution of the target variable P(y) without modification of the conditional relationship P(y|x), affecting class balance.
Ensemble Methods for Drift Detection
Approaches combining multiple drift detectors or base models to improve the robustness and accuracy of change detection.
Gradual Drift
Type of concept drift where changes in data distribution or relationships between variables occur gradually over an extended period.
Abrupt Drift
Sudden and immediate change in data distribution or relationships between variables, requiring rapid detection and model adaptation.