# Prepare training and test samples
olive <- read.csv(file.choose())
head(olive)
dim(olive)

olive.south<-subset(olive, region == "South", select = area:eicosenoic)
head(olive.south)
dim(olive.south)
olive.south$area <- factor(olive.south$area,
     levels=c("Calabria", "North Apulia", "Sicily", "South Apulia"))

plotmatrix(olive.south[,2:9], colour=alpha("black",0.2))
           
p <- ggpcp(olive.south, vars=names(olive.south)[c(2,3,6,4,5,7,8,9)])
p + geom_line(aes(colour=area))

library(rggobi)
ggobi(olive.south)

table(olive.south$area)
table(olive.south$area)*0.67
training.indx<-c(sample(1:56,38), sample(57:81,17), sample(82:117,24),
                 sample(118:323,138))
training.indx<-sort(training.indx)
olive.south.train<-olive.south[training.indx,]
olive.south.test<-olive.south[-training.indx,]

# Trees
library(rpart)
?rpart
olive.tree <- rpart(area ~ ., data=olive.south.train)
table(olive.south.train$area, predict(olive.tree, olive.south.train,
          type="class"))
table(olive.south.test$area, predict(olive.tree, olive.south.test,
          type="class"))

# Random forests
library(randomForest)
?randomForest
olive.rf<-randomForest(area~.,data=olive.south.train,
  importance=TRUE,proximity=TRUE,mtry=2,ntree=1500)
olive.rf$err.rate[1500,]
olive.rf$importance
table(olive.south.train$area,olive.rf$predicted)
table(olive.south.test$area,predict(olive.rf, olive.south.test))
qplot(linoleic, palmitoleic, data=olive.south.train, colour=area)

test.cl <- function(true, pred){
        true <- max.col(true)
        cres <- max.col(pred)
        table(true, cres)
}

# Neural networks
library(nnet)
?nnet

olive.nn<-nnet(area~.,olive.south.train,size=4,linout=T,decay=0.005,
  range=0.06,maxit=1000)
targetr <- class.ind(olive.south.train$type)
targets <- class.ind(olive.south.test$type)
test.cl(targetr, predict(olive.nn,olive.south.train))
test.cl(targets, predict(olive.nn,olive.south.test))
olive.nn.keep <- olive.nn

# Your turn
music <- read.csv(file.choose(), row.names=1)
head(music)
dim(music)
music.train<-music[1:54,-1]
music.test<-music[55:59,-1]

music.rf<-randomForest(type~.,data=music.train,
  importance=TRUE,proximity=TRUE,mtry=2,ntree=500)
music.rf$err.rate[500,]
music.rf$importance
table(music.train$type,music.rf$predicted)
predict(music.rf, music.test)

music.nn<-nnet(type~.,music.train,size=4,decay=0.01,
  maxit=1000)
table(music.train$type, round(predict(music.nn, music.train)))
round(predict(music.nn, music.test))

# Model-based clustering
library(mclust)
help(package="mclust")
data(diabetes)
head(diabetes)
dim(diabetes)

plotmatrix(diabetes[,-1])
ggpcp(diabetes, vars=names(diabetes[,-1])) + geom_line()

diabetes.mc<-mclustBIC(diabetes[,-1])
plot(diabetes.mc)

diabetes.mc.best<-mclustModel(diabetes[,-1], diabetes.mc,
    G=3, modelNames="VVV")
diabetes$cl<-summary(diabetes.mc.best)$classification

#plotmatrix(diabetes[,-1], aes(colour=factor(cl)))
clrs<-brewer.pal(3, "Set2")
pairs(diabetes[,2:4], col=clrs[diabetes$cl],pch=16)
diabetes.mc.best$parameters

# Self-organizing maps
library(kohonen)
help(package="kohonen")
diabetes.som<-som(scale(diabetes[,2:4]), grid=somgrid(5,5, "hexagonal"))
diabetes.som$codes
plot(diabetes.som)

sqrt(sum(diabetes.som$distances^2))

# Show map in data space
f.ggobi.som<-function(x.som) {
  grd <- x.som$grid$pts[x.som$unit.classif,]
  xmx<-jitter(grd[,1],factor=2)
  xmy<-jitter(grd[,2],factor=2)
  ncols <- ncol(x.som$data)
  x.ggobi<-cbind(x.som$data,xmx,xmy)
  dimnames(x.ggobi)[[2]][ncols+1]<-"Map 1"
  dimnames(x.ggobi)[[2]][ncols+2]<-"Map 2"
  x.grid<-cbind(x.som$codes,x.som$grid$pts)
  dimnames(x.grid)[[2]]<-dimnames(x.ggobi)[[2]]
  x.clust<-rbind(x.ggobi,x.grid)
}
f.ggobi.som.net<-function(x.som) {
  x.net<-NULL
  for (i in 1:x.som$grid$xdim) {
    for (j in 1:x.som$grid$ydim) {
        if (j<x.som$grid$ydim) x.net<-rbind(x.net,c((i-1)*x.som$grid$xdim+j,
                       (i-1)*x.som$grid$xdim+j+1))
    if (i<x.som$grid$xdim) x.net<-rbind(x.net,c((i-1)*x.som$grid$xdim+j,
                       i*x.som$grid$xdim+j))
    }
  }
  colnames(x.net)<-c("from","to")
  return(x.net)
}

library(rggobi)
diabetes.som.ggobi<-f.ggobi.som(diabetes.som)
rownames(diabetes.som.ggobi)<-c(1:170)
gg<-ggobi(diabetes.som.ggobi)
g<-gg[1]

diabetes.som.ggobi.net<-f.ggobi.som.net(diabetes.som)
edges(g) <- diabetes.som.ggobi.net + 145

gcolor<- c(rep(9,145),rep(8,25))
glyph_color(g)<-gcolor
gtype<- c(rep(6,145),rep(4,25))
glyph_type(g)<-gtype
gsize<- c(rep(3,145),rep(5,25))
glyph_size(g)<-gsize