| Summary: | 碩士 === 東海大學 === 統計學系 === 91 === Multivariate adaptive regression splines (MARS) is a powerful method to built
a non-linear model by a greedy search procedure.
There are several advantages to use the method such as natural handling of data of mixed type,
computational ability for large sample size, and interpretability.
However, the computational effort is huge when the number of predictors is large
or when the support of each predictor contains many distinct values.
To overcome the pitfalls of the MARS,
we propose a hybrid procedure that uses the linear projected predictors calculated
by Fisher''s discriminate variable or principle component
and proceed the similar procedure of MARS.
In this research, we extend the result by Frideman (2002) and establish the procedure
called ''stochastic boosting projected additive regression splines'''' (SB-PARS).
Boosting is a powerful machine learning.
Stochastic boosting combines randomization into boosting procedure
to increase robustness and the speed of the execution.
We use several data sets to illustrate the benefits of dealing with large data sets
in model building and computational speed.
We compare our proposed procedure with several sophisticated methods,
such as classification and regression trees (CART),
multiple additive regression tree (MART), and stochastic MART.
|
|---|
