Słownik AI
Kompletny słownik sztucznej inteligencji
Min-Max Normalization
Scaling technique that transforms values into a specific range (usually [0, 1]) using the formula (x - min) / (max - min) to preserve relative relationships between observations.
Z-Score Standardization
Centered-reduced transformation that subtracts the mean and divides by the standard deviation, resulting in a distribution with mean 0 and standard deviation 1 to make variables comparable.
Box-Cox Transformation
Parametric transformation that applies an optimized power λ to positive data to stabilize variance and normalize the distribution, particularly effective for right-skewed data.
Yeo-Johnson Transformation
Extension of Box-Cox that handles negative and zero values, using different formulas based on the sign of x to normalize distributions while preserving interpretability.
Logarithmic Transformation
Application of a logarithmic function (usually natural or base 10) to compress scales, reduce skewness, and transform multiplicative relationships into additive ones.
Quantile Transformation
Non-parametric mapping that transforms variables to follow a specific distribution (uniform or Gaussian) using empirical quantiles, robust to extreme values.
Robust Scaler
Scaling method using the median and interquartile range (IQR) instead of the mean and standard deviation, resistant to outliers and non-Gaussian distributions.
Winsorization
Technique for limiting extreme values by replacing observations beyond certain quantiles (typically 1st and 99th percentiles) with these threshold values to reduce the impact of outliers.
Square root transformation
Application of a √x transformation to moderate right skewness, particularly useful for count data or variables following a Poisson distribution.
Unit Vector Scaling
Normalization that divides each vector by its Euclidean norm to obtain unit magnitude, commonly used in text processing and similarity analysis.
Arcsine transformation
Application of arcsin(√x) to proportions or probabilities to stabilize variance and normalize distributions bounded between 0 and 1, particularly in meta-genetics and ecology.
Decimal Scaling
Normalization method that divides values by a power of 10 (10^j) where j is the smallest integer such that max(|x|/10^j) < 1, preserving signs and proportional relationships.
Inverse transformation
Application of the 1/x function to transform distributions with long right tails, particularly effective for variables representing rates or ratios.
Power Transformation
Family of transformations x^λ where λ is an adjustable parameter, encompassing square root (λ=0.5), square (λ=2) and other transformations to model non-linear relationships.
Clipping
Limiting values to a predefined interval by replacing extreme values with upper or lower bounds, a simple alternative to winsorization for outlier treatment.
Adaptive binning
Discretization that creates bins of variable width based on data distribution (quantiles) or statistical criteria, better capturing local structures of continuous data.
Sinusoidal transformation
Application of trigonometric functions to capture cyclical or periodic patterns in temporal data, transforming linear relationships into oscillatory behaviors.