AI-ordlista
Den kompletta ordlistan över AI
Change Detection
Algorithmic process aimed at identifying significant statistical changes in data distribution or relationships between variables. Early detection enables triggering model updates before performance degradation occurs.
Gradual Drift
Type of concept drift where changes in the relationship between variables occur progressively over an extended period. This slow evolution makes detection particularly complex as variations may be mistaken for statistical noise.
Abrupt Drift
Sudden and significant change in data distribution or relationship between variables, occurring over a very short period. Abrupt drifts require rapid detection to avoid major performance degradations of the model.
Recurrent Drift
Phenomenon where concepts or relationships between variables reappear periodically after having temporarily disappeared. Recognition of recurrent patterns enables anticipating changes and optimizing adaptation strategies.
Covariate Drift
Type of drift where the distribution of input variables changes while the conditional relationship P(Y|X) remains stable. This phenomenon particularly affects models sensitive to feature distribution such as Bayesian algorithms.
Prior Drift
Change in the marginal distribution of the target variable P(Y) without modification of the conditional relationship P(Y|X). This drift affects global predictions but preserves local relationships between variables.
Online Adaptation
Process of continuously updating the predictive model as new data arrives, enabling rapid response to concept drifts. Online adaptation balances model stability and reactivity to changes.
Detection Delay
Metric measuring the time elapsed between the actual occurrence of a concept drift and its detection by the algorithm. Minimal delay is crucial to minimize the period of model performance degradation.
False Positive Rate
Proportion of false drift alerts generated by the detection algorithm, leading to unnecessary model updates. Optimizing this rate is essential to maintain stability while ensuring sensitivity.
Supervised Detection
Drift detection method using actual labels to directly analyze changes in the relationship between inputs and outputs. This approach offers high precision but requires continuous availability of ground truth.
Unsupervised Detection
Drift detection approach relying solely on input data features without using labels, typically through analysis of changes in statistical distribution. This method is applicable even when ground truth is not available.