Regression Shrinkage and Selection Via the Lasso-英文文献.pdf

Regression shrinkage and selection via the lasso  Robert Tibshirani y Department of Statistics and Division of Biostatistics Stanford University Abstract We prop ose a new metho d for estimation in linear mo dels The lasso minimizes the residual sum of squares sub ject to the sum of the absolute value of the co ecients b eing less than a constant Because of the nature of this constraint it tends to pro duce some co ecients that are exactly zero and hence gives interpretable mo dels Our simulation studies suggest that the lasso enjoys some of the favourable prop erties of b oth subset selection and ridge regression It pro duces interpretable mo dels like subset selection and exhibits the stability of ridge regression There is also an interesting relationship with recent work in adaptive function estimation by Donoho and Johnstone The lasso idea is quite general and can b e applied in a variety of statistical mo dels extensions to generalized regression mo dels and treebased mo dels are briey describ ed Keywords regression subset selection shrinkage quadratic programming Intro duction Consider the usual regression situation we have data xi y i N i i T where x x x and y are the regressors and resp onse for the ith i ip i observation The ordinary least squares OLS estimates are obtained by mini mizing the residual squared error There are two reasons why the data analyst