KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Out-of-Bag Error
Validation error calculated on observations not selected during bootstrap sampling, providing an unbiased estimate of the model's generalization error. OOB error serves as an alternative to cross-validation in ensemble methods like Random Forest.
Bootstrap Aggregating (Bagging)
Ensemble method that trains multiple models on different bootstrap samples and combines their predictions through majority voting or averaging to reduce variance and improve predictive stability. Bagging is particularly effective for unstable algorithms like decision trees.
Bootstrap Confidence Interval
Confidence interval estimated through bootstrap resampling to quantify the uncertainty of model parameters or performance metrics. Several methods exist: percentile, BCa, and basic bootstrap intervals.
Jackknife Validation
Systematic resampling technique where each observation is omitted one at a time to estimate the bias and variance of an estimator, considered as an approximation of bootstrap. Jackknife is computationally less expensive but less flexible than bootstrap.
Permutation Validation
Non-parametric validation method that evaluates the statistical significance of a model by randomly permuting target labels to create a null performance distribution. This technique allows testing whether the model actually learns meaningful patterns.
Bootstrap Bias Correction
Procedure using bootstrap to estimate and correct the bias of an estimator or model evaluation metric. This correction improves the accuracy of estimates particularly on small samples.
BCa Bootstrap (Bias-Corrected and Accelerated)
Advanced bootstrap method that adjusts confidence intervals by correcting bias and acceleration (skewness) of the bootstrap distribution. BCa provides more accurate intervals than the standard percentile method.
Parametric Bootstrap
Bootstrap variant where samples are generated from a parametric distribution fitted to the data rather than direct resampling. This method is useful when the underlying data model is known or assumed.
Non-parametric Bootstrap
Classical bootstrap approach that resamples directly with replacement from observed data without making assumptions about the underlying distribution. This method is more flexible and robust to violations of distributional assumptions.
Bootstrap Sampling Distribution
Empirical distribution of a statistic obtained by calculating this statistic on multiple bootstrap samples, allowing estimation of the variability and asymptotic properties of the estimator. This distribution serves as the basis for bootstrap inferences.
Bootstrap Standard Error
Estimation of the standard error of a statistic calculated as the standard deviation of its bootstrap distribution, providing a robust measure of precision against distributional assumptions. This estimation is particularly useful for complex statistics.
Bootstrap Hypothesis Testing
Hypothesis testing approach using bootstrap to generate the distribution under the null hypothesis and calculate p-values non-parametrically. This method avoids asymptotic approximations and works with complex test statistics.
Bootstrap Model Selection
Model selection process using bootstrap to evaluate the stability and comparative performance of different candidate models on multiple resamples. This approach quantifies uncertainty in the selection process itself.
Bootstrap Power Analysis
Evaluation of the statistical power of a test or model by simulating multiple bootstrap samples under different hypothetical alternatives. This method allows estimating the required sample size without restrictive distributional assumptions.
Bootstrap Calibration
Technique for adjusting probabilities predicted by a classification model using bootstrap to estimate and correct calibration bias. This method improves the reliability of the model's probability scores.
Bootstrap Ensemble Methods
Family of algorithms combining bootstrap and aggregation to reduce prediction variance, including Random Forest, Extra Trees and other methods based on bootstrap samples. These methods exploit the diversity created by resampling.
Bootstrap Feature Importance
Evaluation of variable importance by measuring their impact on model performance across multiple bootstrap samples, providing confidence intervals for importance. This approach quantifies the stability of feature importance.
Bootstrap Model Stability
Measure of model robustness assessed by the consistency of its predictions or parameters across different bootstrap samples. Stability is crucial for confidence in real-world model deployments.
Bootstrap.632+
Advanced bootstrap variant that corrects the optimistic bias of error estimation by combining bootstrap error and apparent error with a weight of 0.632, adjusted for overfitting. This method provides more accurate error estimates.