A flexible class of parametric models are used for tree splitting using lasso. This includes CART splits, hyperplane, ellipsoid and hyperboloid cuts. Coordinate descent is used for fast calculation of the penalized lasso parametric models. Cross-validation is employed to obtain the lasso regularization parameter. These trees are called super greedy trees (SGTs) and are constructed using best split first (BSF) where cuts are made sequentially in order of greatest empirical risk reduction.
