Efficiently Using Second Order Information in Large l1 Regularization Problems