New Computational and Statistical Aspects of Regularized Regression with Application to Rare Feature Selection and Aggregation