Abstract:Existing grouped variable selection methods rely heavily on prior group information, thus they may not be reliable if an incorrect group assignment is used. In this paper, we propose a family of shrinkage variable selection operators by controlling the k-th largest norm (KAN). The proposed KAN method exhibits some flexible group-wise variable selection naturally even though no correct prior group information is available. We also construct a group KAN shrinkage operator using a composite of KAN constraints. Neither ignoring nor relying completely on prior group information, the group KAN method has the flexibility of controlling within group strength and therefore can reduce the effect caused by incorrect group information. Finally, we investigate an unbiased estimator of the degrees of freedom for (group) KAN estimates in the framework of Stein's unbiased risk estimation. Extensive simulation studies and real data analysis are performed to demonstrate the advantage of KAN and group KAN over the LASSO and group LASSO, respectively.
Keywords: Degrees of freedom | Group shrinkage | k-th largest norm | Shrinkage estimator | Variable selection
Article:Consider a high-dimensional sparse linear regression model, yi=β0+x′iββ+εi, 1≤ i ≤n, (1) where xi=(xi1,…,xip)′ is a p− dimensional predictor, yi is a univariate response variable, and εiεi's are independent and identically distributed random variables. Without loss of the generality, we assume both the response variable and predictors to be centered and standardized such that =1 yi = 0, =1 xij=0 and =1 2 =n. Thus, β0=0 is assumed in the true model. We are interested in estimating regression coefficients vector ββ=(β1,…,βp)′. The true model in (1) is high-dimensional if the number of covariates p is much larger than the sample size n. The model is sparse since most elements in β are zero.Variable selection is an important issue for such a high-dimensional sparse model. In the last twenty years, the least absolute shrinkage selection operator [LASSO, (Tibshirani 1996)] has