Abstract
We propose sparse estimation methods for the generalized linear models, whichrun Least Angle Regression (LARS) and Least Absolute Shrinkage and SelectionOperator (LASSO) in the tangent space of the manifold of the statistical model.Our approach is to roughly approximate the statistical model and tosubsequently use exact calculations. LARS was proposed as an efficientalgorithm for parameter estimation and variable selection for the normal linearmodel. The LARS algorithm is described in terms of Euclidean geometry withregarding correlation as metric of the space. Since the LARS algorithm onlyworks in Euclidean space, we transform a manifold of the statistical model intothe tangent space at the origin. In the generalized linear regression, thistransformation allows us to run the original LARS algorithm for the generalizedlinear models. The proposed methods are efficient and perform well. Real-dataanalysis shows that the proposed methods output similar results as that of the$l_1$-penalized maximum likelihood estimation for the generalized linearmodels. Numerical experiments show that our methods work well and they can bebetter than the $l_1$-penalization for the generalized linear models ingeneralization, parameter estimation, and model selection.