KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Marginal Effect
The expected change in the model's prediction when modifying one feature while keeping the distribution of other features constant.
Centered ICE Plot
A variant of the ICE plot where each individual curve is centered on a reference point, facilitating the comparison of relative effects between observations.
Derivative ICE (d-ICE)
A visualization that plots the local derivative of ICE curves, highlighting points where the effect of a feature changes most significantly.
Heterogeneity Analysis
The study of the variation of a feature's effects on predictions across different observations, revealed by ICE curves as opposed to the single PDP curve.
Monte Carlo Integration
A numerical estimation method used to calculate expected values in PDPs by randomly sampling from the distribution of marginalized features.
Grid Approximation
A technique for calculating PDPs by discretizing the feature space on a regular grid and evaluating the model at each grid point.
Feature Correlation Bias
The bias introduced in PDPs when features are correlated, because marginalization creates unrealistic feature combinations in the marginal distribution.
Variance-Based Sensitivity Analysis
A quantitative method that decomposes the variance of predictions into contributions from each feature, complementary to the qualitative information from PDPs.
Functional ANOVA Decomposition
A theoretical framework that decomposes the prediction function into main effects and interactions, mathematically justifying PDPs as estimators of main effects.
Bootstrap Confidence Intervals
A technique for quantifying the uncertainty of PDP estimates by resampling data and calculating confidence intervals for the dependence curves.
Rug Plot
A complementary visualization added to PDPs showing the distribution of feature values along the x-axis, indicating regions of low or high data density.
Prediction Function Decomposition
The mathematical process of separating the complex prediction function into interpretable components, where PDPs represent the effects of individual features.