lasso: a standalone construct

The lasso (least absolute shrinkage and selection operator) is a regression method that adds a penalty on the sum of the absolute values of the coefficients, an L1 penalty, to the usual least-squares fit. Because that penalty has corners at zero, it shrinks many coefficients exactly to zero, so the lasso fits the model and selects variables at once, yielding a sparse, interpretable result. Tibshirani's 1996 Regression shrinkage and selection via the lasso introduced it. The referent is this specific L1-penalised estimator, distinct from the smooth L2 shrinkage of ridge regression, which shrinks coefficients toward but not to zero.[1]

Unit

The size of the L1 penalty controlling how many coefficients are driven to zero

Recovery means

Not applicable

Measured by

Tuned by the penalty parameter via cross-validation, trading prediction error against the number of nonzero coefficients