AI-ordlista
Den kompletta ordlistan över AI
Pool Adjacent Violators Algorithm
Optimized algorithm for solving the isotonic regression problem by iteratively merging adjacent pairs that violate the monotonicity constraint, guaranteeing a quasi-linear time solution.
Piecewise-Constant Function
Mathematical function that maintains a constant value over defined intervals, used in isotonic regression to simplify the calibration function while respecting monotonicity.
Monotonic Constraint
Constraint imposed during optimization guaranteeing that the calibration function is strictly or weakly increasing, preserving the relative order of original predictions.
Non-parametric Calibration
Calibration approach making no assumptions about the underlying data distribution or the parametric form of the transformation function, offering maximum flexibility.
Confident Learning
Learning paradigm that automatically identifies and corrects labeling errors by using prediction confidence estimates to distinguish noisy examples from real patterns.
Calibration Map
Matrix representation of the calibration function showing how each input probability is transformed into a calibrated output probability, visualizing local adjustments.
Rank Consistency
Property guaranteed by isotonic regression that preserves the relative order of original predictions after calibration, essential for maintaining ranking decisions.
Isotonic Constraint
Mathematical restriction requiring that the calibration function preserves the monotonic increasing order, fundamental for guaranteeing semantic consistency of transformed probabilities.
V-Platt Scaling
Variant of Platt Scaling using cross-validation to avoid overfitting, particularly important when the calibration set is of limited size.
Calibration Curve
Empirical curve plotting the relationship between predicted probabilities and actual occurrence rates, serving as a visual basis for evaluating and comparing different calibration methods.