Statements (50)
Predicate | Object |
---|---|
gptkbp:instanceOf |
gptkb:model
statistical analysis regression analysis method |
gptkbp:advantage |
can be biased for large coefficients
not ideal when all predictors are relevant performs variable selection produces sparse models reduces model complexity |
gptkbp:alsoKnownAs |
gptkb:Least_Absolute_Shrinkage_and_Selection_Operator
|
gptkbp:appliesTo |
gptkb:signal_processing
finance genomics bioinformatics image analysis |
gptkbp:assumes |
linear relationship between predictors and response
|
gptkbp:canBe |
high-dimensional data
multicollinearity |
gptkbp:canSetCoefficientsToZero |
yes
|
gptkbp:category |
supervised learning
sparse modeling |
gptkbp:contrastsWith |
gptkb:Ridge_regression
|
gptkbp:differenceFromRidge |
Lasso uses L1 penalty, Ridge uses L2 penalty
|
gptkbp:form |
minimize (1/2n)||y - Xβ||^2_2 + λ||β||_1
|
https://www.w3.org/2000/01/rdf-schema#label |
Lasso regression
|
gptkbp:implementedIn |
gptkb:SAS
gptkb:MATLAB gptkb:scikit-learn R |
gptkbp:introduced |
gptkb:Robert_Tibshirani
|
gptkbp:introducedIn |
1996
|
gptkbp:lambdaParameter |
controls strength of regularization
|
gptkbp:limitation |
can select at most n variables if n < p
can be unstable when predictors are highly correlated |
gptkbp:objective |
minimize sum of squared errors plus lambda times sum of absolute values of coefficients
|
gptkbp:penalty |
L1 norm
|
gptkbp:relatedTo |
gptkb:Elastic_Net
gptkb:Ridge_regression Subset selection |
gptkbp:shrinksCoefficients |
yes
|
gptkbp:solvedBy |
gptkb:least_angle_regression_(LARS)
coordinate descent proximal gradient methods subgradient methods |
gptkbp:usedFor |
feature selection
regularization preventing overfitting |
gptkbp:usedIn |
gptkb:generalized_linear_models
linear regression |
gptkbp:bfsParent |
gptkb:Scikit-learn
|
gptkbp:bfsLayer |
5
|