ConfidenceIntervalMethod

Enums · AiDotNet.Evaluation.Enums

Specifies the method for computing confidence intervals on evaluation metrics.

For Beginners

When you calculate a metric like accuracy = 85%, that’s just a point estimate. The “true” accuracy might be anywhere from 82% to 88%. A confidence interval tells you this range. Different methods compute this range differently:

  • Bootstrap methods: Resample your data many times to see how the metric varies
  • Analytical methods: Use mathematical formulas (faster but more assumptions)

Bootstrap methods are generally preferred as they make fewer assumptions about your data.

How It Works

Confidence intervals quantify the uncertainty in metric estimates. Different methods make different assumptions about the underlying distribution and have varying accuracy.

Fields

Field Summary
ConfidenceIntervalMethod AgrestiCoull Agresti-Coull adjusted interval: Adds pseudo-observations for better coverage.
ConfidenceIntervalMethod BCaBootstrap Bias-Corrected and Accelerated (BCa) bootstrap: Corrects for bias and skewness.
ConfidenceIntervalMethod BasicBootstrap Basic bootstrap (reverse percentile): Reflects the bootstrap distribution.
ConfidenceIntervalMethod BayesianCredible Bayesian credible interval: Uses posterior distribution from Bayesian inference.
ConfidenceIntervalMethod ClopperPearsonExact Clopper-Pearson exact interval: Conservative exact interval for proportions.
ConfidenceIntervalMethod DeLong DeLong method: Specialized method for AUROC confidence intervals.
ConfidenceIntervalMethod NormalApproximation Normal approximation: Simple z-interval using normal distribution.
ConfidenceIntervalMethod PercentileBootstrap Percentile bootstrap: Uses percentiles of the bootstrap distribution directly.
ConfidenceIntervalMethod StudentizedBootstrap Studentized (bootstrap-t) bootstrap: Uses t-statistics for pivoting.
ConfidenceIntervalMethod TDistribution t-distribution interval: Uses t-distribution instead of normal.
ConfidenceIntervalMethod WilsonScore Wilson score interval: Analytical method for proportions/accuracy.
int value__