In Sections 3.2 and 3.3 the MARS procedure was motivated as a series of conceptually simple extensions to recursive partitioning regression. In terms of implementation, however, these ex- tensions produce a dramatic change in the algorithm. The usual implementations of recursive partitioning regression [AID (Morgan and Sonquist, 1963) and CART (Breiman, et al., 1984)] take strong advantage of the special nature of step functions, along with the fact that the result- ing basis functions have disjoint support, to dramatically reduce the computation associated with the middle and inner For-loops of Algorithm 1 (lines 4 and 5). In the case of least-squares fitting, very simple updating formulae can be employed to reduce the computation for the associated lin- ear (least-squares) fit (line 7) from O(NM2 + M3) to O(1).
đang được dịch, vui lòng đợi..
