Title: Penalized angular regression for personalized predictions
Authors: Kristoffer Hellton - Norwegian Computing Center (Norway) [presenting]
Abstract: A novel penalized regression method is introduced which is specifically constructed to personalize predictions. Personalized angle (PAN) regression estimates a covariate vector-specific regression coefficients, shrinking them in terms of their angles utilizing a hyperspherical parametrization. It is shown that the PAN estimate will be the solution of a low-dimensional eigenvector problem, which for an orthonormal design matrix has an explicit solution. We prove that by combining the PAN and the L2 penalty the resulting prediction will have uniformly smaller asymptotic mean squared error than ridge regression. The resulting estimator is illustrated in a medical application.