Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Feature Selection
Process of automatically selecting the most relevant variables to build an optimal predictive model, reducing dimensionality and improving generalization.
Filter Methods
Feature selection techniques independent of the model, evaluating each variable individually according to statistical criteria before training.
Wrapper Methods
Selection approaches using the predictive model to evaluate feature subsets, often more accurate but computationally intensive.
Embedded Methods
Strategies combining selection and learning, where the selection process is directly integrated into the model training algorithm.
Recursive Feature Elimination
Iterative algorithm progressively removing the least important features by retraining the model at each step until reaching the optimal number of variables.
Mutual Information
Measure quantifying the statistical dependence between two variables, used to evaluate feature relevance relative to the target variable.
Variance Threshold
Basic filtering technique eliminating features with variance below a predefined threshold, considered uninformative.
Chi-square Test
Statistical test evaluating independence between categorical variables, used to measure the relevance of qualitative features relative to the target.
ANOVA F-Test
Statistical test comparing variances between groups to evaluate the relationship between numerical features and categorical target variables.
Correlation Coefficient
Statistical measure quantifying the strength and direction of the linear relationship between two variables, used to detect multicollinearity.
Sequential Selection
Greedy method sequentially adding (forward) or removing (backward) features to optimize a model performance metric.
Boruta Algorithm
Wrapper method based on random forests identifying all relevant features by comparing their importance to random shadow variables.
Permutation Importance
Model-agnostic technique evaluating feature importance by measuring performance degradation after random permutation of their values.
Relief Algorithm
Filter method assessing feature relevance by measuring their ability to distinguish neighboring instances of different classes.