Parallel MARS algorithm based on B-splines (Q1424604)

The authors introduce a modification of \textit{J. H. Friedman}'s multivariate adaptive regression splines (MARS) algorithm [Ann. Stat., 19, 1--141 (1991; Zbl 0765.65064)] designed for flexible modelling of high-dimensional data. A new version of MARS called BMARS uses B-splines instead of truncated power basis functions. A modified forward stepwise procedure of BMARS is described as well as backward elimination procedure. In addition, rather simple but efficient parallel version of BMARS is developed. Results of BMARS applications to simulated and real data sets demonstrated its computational efficiency and ability to generate models competitive with those of original MARS.