scalabl ahead linear regress tree david vogel ognian asparouhov tobia scheffer decis tree algorithm base split decis piecewis constant model split algorithm extrapol tree constant model leaf node motiv ahead linear regress tree llrt method propos date scalabl approach exhaust evalu model leaf node order obtain optim split optim llrt gener evalu thousand linear regress model exhaust evalu split node base qualiti fit linear regress model result branch decompos calcul residu sum squar larg part pre comput result method highli scalabl observ obtain high predict accuraci problem strong mutual depend attribut report experi simul real data set 