Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Leave-One-Out Cross-Validation
Cross-validation technique where each observation in the dataset is used once as a test set, with the remaining N-1 observations used for training. This method maximizes data usage but results in high computational complexity.
Mean Squared Error
Evaluation metric measuring the average of the squares of the differences between predicted values and actual values, particularly used in regression problems. It penalizes large errors more heavily due to the quadratic effect.
Bias-Variance Tradeoff
Fundamental dilemma in machine learning where reducing bias increases variance and vice versa, affecting the model's generalization ability. Optimization involves finding the ideal balance between these two components of error.
Generalization Error
Measure of a model's performance on unseen data not used during training, reflecting its ability to generalize. It is estimated through cross-validation techniques to avoid the biased optimism of training error.
Underfitting
Situation where a model is too simple to capture the underlying structure of the data, resulting in poor performance on both training and test sets. It manifests as high and systematic errors during cross-validation.
Learning Curve
Graph representing the evolution of model performance as a function of training set size, revealing overfitting or underfitting problems. It helps determine whether adding more data could improve performance.
Nested Cross-Validation
Advanced technique using two loops of cross-validation for hyperparameter selection and model evaluation, preventing information leakage. The inner loop optimizes hyperparameters while the outer loop evaluates final performance.
Confidence Interval
Estimated range of values containing the true value of a parameter with a specified confidence level, calculated from cross-validation results. It quantifies the uncertainty associated with model performance estimates.
Statistical Significance Testing
Procedure determining whether performance differences between models are statistically significant or due to chance. Tests such as paired Student's t-test or Wilcoxon test are applied to cross-validation scores.
Leave-P-Out Cross-Validation
Generalization of LOOCV where P observations are left out for testing at each iteration, creating C(N,P) possible combinations. It offers a compromise between LOOCV and K-fold in terms of bias-variance and computational cost.
Repeated Cross-Validation
Technique repeating K-fold cross-validation multiple times with different initial partitions to reduce the variance of performance estimation. It provides a more stable evaluation at the cost of increased computation time.
Hyperparameter Tuning
Process of optimizing model hyperparameters using cross-validation to evaluate each configuration and avoid overfitting on the test set. Approaches include Grid Search, Random Search, and Bayesian Optimization.