Clustering

getwd()

setwd("C:/Users/ssn/Documents/Dinesh Analyst/R")

require(xlsx)

seed = read.xlsx("seed.xlsx",sheetName="Sheet1")

seed_train = seed[,c(-8)]

seed_train$Type= NULL

corrplot(cor(seed_train), order = "hclust", addrect = 3)

wssplot <- function(data, nc=15, seed=1234){

        wss <- (nrow(data)-1)*sum(apply(data,2,var))

        for (i in 2:nc){

                set.seed(seed)

                wss[i] <- sum(kmeans(data, centers=i)$withinss)}

        plot(1:nc, wss, type="b", xlab="Number of Clusters (k)",

             ylab="Within groups sum of squares", main="Optimal k with wss")}

wssplot(seed_train)

library("devtools")

install_github("kassambara/factoextra")

library("factoextra")

set.seed(123)

fviz_nbclust(seed_train, kmeans, method = "wss") +

        geom_vline(xintercept = 3, linetype = 2)

### Hierarchical clustering

seed_train$Type= NULL

seed.dist = dist(seed_train)

seed.hclust = hclust(seed.dist)

plot(seed.hclust,main='Default from hclust')

rect.hclust(seed.hclust,3)

groups.3 = cutree(seed.hclust,k=3)

table(groups.3)

table(groups.3, seed$Types)

seed[groups.3 == 2,]

library(fpc)

plotcluster(seed_train,groups.3)

table(groups.3,seed$Types)

groups.3  1  2  3

       1 52 23  0

       2 18  0 71

       3  0 47  0

###PAM: Partitioning around Medoids

library(NbClust)

library(cluster)

seed.pam = pam(scale(seed_train),3,FALSE,"euclidean")

seed.pam$medoids

clusplot(seed.pam, main = "Cluster plot, k = 3",color = TRUE)

fviz_cluster(seed.pam)

table(seed$Types,seed.pam$clustering)

fviz_silhouette(silhouette(seed.pam))

## It can be seen that some samples have a negative silhouette. This means that they are not in the right cluster

#We can find the name of these samples and determine the clusters they are closer, as follow:

cluster size ave.sil.width

1       1   71          0.33

2       2   62          0.49

3       3   78          0.39

# Compute silhouette

sil <- silhouette(seed.pam)[, 1:3]

# Objects with negative silhouette

neg_sil_index <- which(sil[, 'sil_width'] < 0)

sil[neg_sil_index, , drop = FALSE]

#######

#### k means cluster analysis

#http://beyondvalence.blogspot.in/2013/12/cluster-analysis-choosing-optimal.html

####determine no. of clusters

library(NbClust)

set.seed(1234)

nc <- NbClust(seed_train, min.nc=2, max.nc=15, method="kmeans")

table(nc$Best.n[1,])

barplot(table(nc$Best.n[1,]),

        xlab="Numer of Clusters", ylab="Number of Criteria",

        main="Number of Clusters Chosen by 26 Criteria")

set.seed(1234)

fit.km <- kmeans(seed_train, 3)                          

fit.km$size

fit.km$centers

aggregate(seed_train, by=list(cluster=fit.km$cluster), mean)

sil<-silhouette(fit.km$cluster,seed.dist)

plot(sil,cex.names=0.6,nmax=98,main="Silhouette Plot, k=2", col=c("red","green"))

plot(seed_train, col = fit.km$cluster, pch = 19, frame = FALSE,

     main = "K-means with k = 3")

points(fit.km$centers, col = 1:4, pch = 8, cex = 3)

fviz_cluster(fit.km, data = seed_train)

library(flexclust)

#We can check how well the kmeans clusters =3 clustered the data by using the flexclust package.

#See how well the clustering was for channel 1 and 2? Not very well with a low index of 0.152.

0  2  3  4  7 11 12 15  2  7 10  3  1  1  1  1

>

#library(fpc)

w.table<-table(seed$Types,fit.km$cluster)

w.table

randIndex(w.table)

##############

library(clValid)

intern <- clValid(seed_train, 2:8, clMethods=c("hierarchical","kmeans","pam"),validation=c("internal","stability"))

res <- getRanksWeights(intern)

print(res$ranks[,1:3], quote=FALSE)

optimalScores(intern)

p <- par(no.readonly=TRUE)

par(mfrow=c(2,2),mar=c(4,4,3,1))

plot(intern, legend=FALSE)

par(op)

plot(nClusters(intern),measures(intern,"Dunn")[,,1],type="n",axes=F, xlab="",ylab="")

legend("center", clusterMethods(intern), col=1:9, lty=1:9, pch=paste(1:9))

> print(res$ranks[,1:3], quote=FALSE)

             1              2              3            

APN          pam-3          kmeans-3       pam-2        

AD           pam-8          kmeans-8       kmeans-7     

ADM          pam-3          kmeans-3       pam-6        

FOM          hierarchical-8 kmeans-8       hierarchical-7

Connectivity hierarchical-2 pam-2          kmeans-2     

Dunn         hierarchical-8 hierarchical-7 hierarchical-6

Silhouette   hierarchical-2 kmeans-2       pam-2        

> optimalScores(intern)                   Score       Method Clusters APN          0.05254478          pam        3 AD           1.54480205          pam        8 ADM          0.21938308          pam        3 FOM          0.46854722 hierarchical        8 Connectivity 8.78611111 hierarchical        2 Dunn         0.11986747 hierarchical        8 Silhouette   0.52534203 hierarchical        2