Shrinkage Estimators for the Intercept in Linear and Uplift Regression

Shrinkage estimators modify classical statistical estimators by scaling them towards zero in order to decrease their prediction error. We propose shrinkage estimators for linear regression models which explicitly take into account the presence of the intercept term, shrinking it independently from other coefficients. This is different from current shrinkage estimators, which treat the intercept just as an ordinary regression coefficient. We demonstrate that the proposed approach brings systematic improvements in prediction accuracy if the true intercept term differs in magnitude from other coefficients, which is often the case in practice. We then generalize the approach to uplift regression which aims to predict the causal effect of a specific action on an individual with given characteristics. In this case the proposed estimators improve prediction accuracy over previously proposed shrinkage estimators and achieve impressive performance gains over original models.