Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Robust Matrix Factorization
Set of matrix decomposition techniques designed to resist outliers and noise, minimizing the influence of data errors on estimated latent factors.
Robust Singular Value Decomposition (rSVD)
Variant of standard SVD that uses outlier-resistant norms (such as L1 norm) instead of L2 norm, enabling more reliable estimation of principal components in the presence of noise.
Robust Non-Negative Matrix Factorization (rNMF)
Extension of NMF that incorporates mechanisms for detecting and handling outliers, often through regularization terms or specific loss functions, to ensure non-negative and stable factors.
Huber Method in Factorization
Approach that applies the Huber loss function, quadratic for small errors and linear for large ones, to factorization optimization, offering a compromise between noise sensitivity and outlier robustness.
L1-PCA Factorization
Principal component analysis method based on minimizing the L1 norm, which is less sensitive to extreme points than the L2 norm traditionally used in standard PCA.
Robust Principal Component Analysis (RPCA)
Technique that decomposes a data matrix into the sum of a low-rank matrix (underlying structure) and a sparse matrix (outliers), thus effectively isolating signal from noise.
Median Absolute Deviation Matrix Factorization (MAD)
Robust method that uses median absolute deviation as a dispersion measure to weight errors during decomposition, reducing the impact of extreme observations on the final model.
Robust Alternating Least Squares Algorithm
Variant of the ALS algorithm where the loss function is modified to be less sensitive to outliers, often by replacing least squares with robust criteria during alternating factor updates.
Robust Quantile Factorization
Approach that optimizes factorization based on quantiles of residuals rather than their mean, making the estimation process insensitive to a certain proportion of corrupted data.
Robust Latent Factor Decomposition
Modeling that extracts latent variables (factors) from a data matrix while mitigating the effect of noisy or outlier observations, for a more faithful representation of the intrinsic structure of the data.
Matrix Factorization with Robustness Weights
Technique that assigns a weight to each matrix entry based on its probability of being an outlier value, with weights then used to minimize the influence of unreliable data during decomposition.
RANSAC Method in Factorization
Application of Random Sample Consensus (RANSAC) to factorization, where the model is iteratively estimated from random subsets of data, retaining only inliers (consistent data) to build the final robust model.
L2,1 Norm Factorization
Method that minimizes the L2,1 norm of the error matrix (sum of L2 norms of rows), making it particularly robust to structured outliers in rows or columns of the data matrix.
Robust Truncated Decomposition (Robust Truncated SVD)
Robust version of truncated SVD that, before retaining the k largest singular values, applies a cleaning or weighting procedure to mitigate the effect of noise and extreme points on the retained components.
Matrix Factorization with Noise Mixture Model
Advanced approach that explicitly models noise as a mixture of distributions (e.g., a Gaussian distribution for noise and a wider one for outliers), allowing for finer separation and more stable factorization.