Josh Day
February 5, 2014
adabag
package in R
[2]
randomForest
package in R
[2]
[2]
[2]
Consider estimating real-valued function
\[ f^0(\cdot) = \text{arg}\,\text{min}_{f(\cdot)} \; \text{E}[\;\rho(f(X), Y)\;] \]
where \( \rho(\cdot) \) is a loss function, typically assumed differentiable and convex with respect to the first argument.
Minimization is over all (measurable) functions \( f(\cdot) \)
\[ f^0(\cdot) = \text{arg}\,\text{min}_{f(\cdot)} \; \text{E}[\;\rho(f(X), Y)\;] \]
Estimation of \( f^0(\cdot) \) with boosting can be done by using empirical risk: \[ \frac{1}{n}\sum_{i=1}^n \rho(f(X_i), Y_i) \]
and pursuing iterative steepest descent in function space.
[1]
[2]
FGD algorithm using squared error loss function: \[ \rho_{L_2}(y,f) = \frac{(y-f)^2}{2} \]
Very useful for high-dimensional regression
Simple, fast
[1]
gbm
packagelibrary(gbm)
Boston.boost <- gbm(medv~., data=Boston, distribution='gaussian', n.trees=5000, train.fraction=.7)
'gaussian'
option uses squared error loss
summary(Boston.boost, plot=FALSE)
var rel.inf
rm rm 80.360922
lstat lstat 16.386085
ptratio ptratio 1.408125
tax tax 1.376013
dis dis 0.277621
indus indus 0.067403
crim crim 0.051962
age age 0.037138
black black 0.025459
nox nox 0.005013
rad rad 0.003118
chas chas 0.001139
zn zn 0.000000