# Emulating example 14.3.8 in HTF: K-means clustering wtih Human Tumor Microarray Data
# Jill McCracken
# September 18, 2009

cancerdat <- read.csv ("C:/Documents and Settings/m29287/My Documents/Analytics/Statistics/HTF -- Summer 2009/Clustering example/cancer_dataT.csv", header = TRUE)

dim(cancerdat)

# cancerdat contains both the train / test data sets from HTF 
# It has been TRANSPOSED prior to loading into R
# note that cancerdat has 198 rows (cancer samples) and 16066 columns
# the first three columns indicate cancer type and sample from HTF website
# Columns 4 - 16066 are the different genes, our clustering variables

cancerdatall <- cancerdat[,4:16066]
types <- cancerdat[,1:3]

# scale the data prior to clustering

scale(cancerdatall)

# set the seed so we can reproduce results (so kmeans will pick the same initial cluster centers)
set.seed(123)

# start with clustering all of the data, k = 4

cl <- kmeans(cancerdatall,4)
ssall <- sum(cl$withinss)
ssall

# now try clustering all of the data with k = 2 through 9

clk <- matrix(0,9,2)

for (i in 1:9)
{
	clkclus <- kmeans(cancerdatall,i+1)
	clk[i,1] <- i
	clk[i,2] <- sum(clkclus$withinss)
}

clres <- cbind(types,cl$cluster)
table(clres$type, clres[,4])

# split the data into three equal sized groups of genes and compare results for k = 3

cancerdat1 <- cancerdat[,4:5357]
cancerdat2 <- cancerdat[,5358:10712]
cancerdat3 <- cancerdat[,10713:16066]
scale(cancerdat1)
scale(cancerdat2)
scale(cancerdat3)

cl1 <- kmeans(cancerdat1, 3)
cl2 <- kmeans(cancerdat2, 3)
cl3 <- kmeans(cancerdat3, 3)

ss1 = sum(cl1$withinss)
ss2 = sum(cl2$withinss)
ss3 = sum(cl3$withinss)

ss1
ss2
ss3


cl1$size
cl2$size
cl3$size

cl1res <- cbind(types,cl1$cluster)
cl2res <- cbind(types,cl2$cluster)
cl3res <- cbind(types,cl3$cluster)

table(cl1res$type,cl1res[,4])
table(cl2res$type,cl2res[,4])
table(cl3res$type,cl3res[,4])

# cancerdat3 has the lowest WithinSS and good separation of the cancer types, so let's try that data set for different values of k

cl3k <- matrix(0,9,2)

for (i in 1:9)
{
	cl3kclus <- kmeans(cancerdat3,i+1)
	cl3k[i,1] <- i+1
	cl3k[i,2] <- sum(cl3kclus$withinss)
}

plot(cl3k[,1],cl3k[,2],xlab="k",ylab="TotWithinSS",col="red")


cl3k8clus <- kmeans(cancerdat3,8)
cl3k8clus$size
cl3k8res <- cbind(types,cl3k8clus$cluster)
table(cl3k8res$type,cl3k8res[,4])


# Look at results for 14 clusters

cl3k14clus <- kmeans(cancerdatall,14)
cl3k14clus$size

cl3k14res <- cbind(types,cl3k14clus$cluster)
table(cl3k14res$type,cl3k14res[,4])