## source("da.R")

## bank data ################################################

library(Rfwdmv)
data(bank.dat)

x <- bank.dat

x <- data.frame(x,c(rep("genuine",100),rep("forged",100)))
names(x) <- c("length","left.height","right.height",
              "lower.frame","upper.frame","diagonal","group")

train <- sort(sample(1:nrow(x),nrow(x)*0.75))
test  <- (1:nrow(x))[-train]

sort(c(train,test)) ## check

table(x$group[train])
table(x$group[test])

## LDA ######################################################

l1 <- lda(group ~ diagonal, data=x[train,])
plot(l1)

pl1 <- predict(l1,newdata=x[test,]) ## predict with test
plot((test),pl1$x,col="white")
text((test),pl1$x,substr(as.character(x$group[test]),1,1))
table(pl1$class,x$group[test])

pl1 <- predict(l1,newdata=x[train,]) ## predict with train
plot((train),pl1$x,col="white")
text((train),pl1$x,substr(as.character(x$group[train]),1,1))
table(pl1$class,x$group[train])

l2 <- lda(group ~ lower.frame + upper.frame + diagonal, data=x[train,])
plot(l2)

pl2 <- predict(l2,newdata=x[test,]) ## predict with test
plot((test),pl2$x,col="white")
text((test),pl2$x,substr(as.character(x$group[test]),1,1))
table(pl2$class,x$group[test])

pl2 <- predict(l1,newdata=x[train,]) ## predict with train
plot((train),pl2$x,col="white")
text((train),pl2$x,substr(as.character(x$group[train]),1,1))
table(pl2$class,x$group[train])

l3 <- lda(group ~ length + left.height + right.height + lower.frame + upper.frame + diagonal, data=x[train,])
plot(l3)

pl3 <- predict(l3,newdata=x[test,]) ## predict with test
plot((test),pl3$x,col="white")
text((test),pl3$x,substr(as.character(x$group[test]),1,1))
table(pl3$class,x$group[test])

pl3 <- predict(l1,newdata=x[train,]) ## predict with train
plot((train),pl3$x,col="white")
text((train),pl3$x,substr(as.character(x$group[train]),1,1))
table(pl3$class,x$group[train])

## QDA ######################################################

q3 <- qda(group ~ length + left.height + right.height + lower.frame + upper.frame + diagonal, data=x[train,])
##plot(q3)   ## plot does not work for q3, but rest is OK

pq3 <- predict(q3,newdata=x[test,]) ## predict with test
table(pq3$class,x$group[test])

pq3 <- predict(l1,newdata=x[train,]) ## predict with train
table(pq3$class,x$group[train])

## Logit ####################################################

lg1 <- glm(group ~ length + left.height + right.height + lower.frame + upper.frame, data=x[train,], family=binomial())
summary(lg1)

p1 <- predict(lg1,newdata=x[test,])
group.p1 <- 1/(1+exp(-p1)) > 0.5
table(group.p1, x$group[test])

p1 <- predict(lg1,newdata=x[train,])
group.p1 <- 1/(1+exp(-p1)) > 0.5
table(group.p1, x$group[train])

## may not not work with all variables since perfect separation possible!
lg2 <- glm(group ~ length + left.height + right.height + lower.frame + upper.frame + diagonal, data=x[train,], family=binomial())  ## -> no convergence
summary(lg2)

p2 <- predict(lg2,newdata=x[test,])
group.p2 <- 1/(1+exp(-p2)) > 0.5
table(group.p2, x$group[test])

p2 <- predict(lg2,newdata=x[train,])
group.p2 <- 1/(1+exp(-p2)) > 0.5
table(group.p2, x$group[train])