library(MASS)
library(mda)

# MARS function from MDA library does NOT use any of Friedman's MARS code
# Therefore the results are similar to Friedman's but NOT the same.
# This function does NOT handle properly CATEGORICAL predictors; 
# Is NOT well-tested 

# Among the arguments of MARS function
# the maximum number of model terms: nk = max(21, 2 * ncol(x) + 1)
# degree - optional integer specifying the maximum interaction degree

# =========================================================================

spam<-read.table("C:\\AnaMaria\\DATA\\spam.data", sep="")

# use only a fragment of data (size = 3000) 
# use a test set of size 1000 
# use a train set of size  2000 

spam.red <- spam[sample(1:4601,3000),]
rand<-sample(1:3000,1000)
test.spam<-spam.red[rand,]
train.spam<-spam.red[-rand,]
X.spam <- train.spam[,1:57]
y.spam <- train.spam[,58]
train.MARS <- mars(X.spam, y.spam, degree=2)

# train.MARS$gcv
# [1] 0.05910394

X.spam.test<-test.spam[,1:57]
y.spam.test<-test.spam[,58]

missclassif <- matrix(0,1000,100)
error <- rep(0,100)
for (i in 1:100)
{
	train.MARS<- mars(X.spam, y.spam, degree=2, nk=i)
	missclassif[,i]<-predict.mars(train.MARS, X.spam.test)
	error[i] <- sum((y.spam.test-missclassif[,i] )^2)/1000
}

rank <- c(1:100)
plot(x=rank, y=error, xlab="Rank of Model", ylab="Test Misclassification Error", xlim=c(0,100), ylim=c(0.05,0.4),cex=0.5, col="green", pch=7)
lines(x=rank.red, y=err.red, lty=1, col="green")
abline(h=0.055, lty=3)