Supervised Learning -a.Rmd

---
title: "Supervised learning - Part1"
author: "Yatish"
date: "October 14, 2015"
output: html_document
---
### Taking input of leaf data.

```{r}
#install.packages("class")
library("class")
lf<- read.csv("leaf.csv")
colnames(lf)<- c("Class", "SpecimenNumber","Eccentricity","AspectRatio","Elongation","Solidity","StochasticConvexity","IsoperimetricFactor","MaximalIndentationDepth","Lobedness","AverageIntensity","AverageContrast","Smoothness","ThirdMoment","Uniformity","Entropy")

table(lf$Class)
lf$Class
```

Shuffling the rows to rearrange the data
```{r}
shuff<-runif(nrow(lf))
leaf<-lf[order(shuff),]
leaf$Class
```

Scaling the data.
```{r}
leaf.scaled<-as.data.frame(lapply(leaf[,c(2:16)], scale))
head(leaf.scaled)
summary(leaf.scaled)

```

Normalizing the data
```{r}
normalize<- function(x) {
  return((x-min(x))/(max(x)-min(x)))
}
leaf.normalized<-as.data.frame(lapply(leaf[,c(2:16)],normalize))
head(leaf.normalized)
```

Classifying the data using k-NN.
```{r}
leaf.normalized.train<-leaf.normalized[1:300,]
leaf.normalized.test<-leaf.normalized[301:339,]
leaf.normalized.train.target<-leaf[1:300,c(1)]
leaf.normalized.test.target<-leaf[301:339,c(1)]

k<-5
knn.m1<- knn(train=leaf.normalized.train, test=leaf.normalized.test,leaf.normalized.train.target,k)
length(knn.m1)

cm<-table(leaf.normalized.test.target,knn.m1)
cm
```

Changing k values.
```{r}
  k<-8
knn.m1<- knn(train=leaf.normalized.train, test=leaf.normalized.test,leaf.normalized.train.target,k)
length(knn.m1)

cm<-table(leaf.normalized.test.target,knn.m1)
cm
k<-2
knn.m1<- knn(train=leaf.normalized.train, test=leaf.normalized.test,leaf.normalized.train.target,k)
length(knn.m1)

cm<-table(leaf.normalized.test.target,knn.m1)
cm
```

By looking at the above data we can clearly state that changing the K values change the data drastically. With K=5 there were two observations in leaf 1(Quercus suber) whereas changing the K values to 8 changed the cluster position and k value to 2 came back with a tie breaker thus less efficient.


Using k-NN with scaled and unscaled data
```{r}
#scaled data
leaf.scaled.train<-leaf.scaled[1:300,]
leaf.scaled.test<-leaf.scaled[301:339,]
leaf.scaled.train.target<-leaf[1:300,c(1)]
leaf.scaled.test.target<-leaf[301:339,c(1)]

k<-5
knn.m1<- knn(train=leaf.scaled.train, test=leaf.scaled.test,leaf.scaled.train.target,k)
length(knn.m1)

cm<-table(leaf.scaled.test.target,knn.m1)
cm

#raw data
leaf.data.train<-leaf[1:300,]
leaf.data.test<-leaf[301:339,]
leaf.data.train.target<-leaf[1:300,c(1)]
leaf.data.test.target<-leaf[301:339,c(1)]

k<-5
knn.m1<- knn(train=leaf.data.train, test=leaf.data.test,leaf.data.train.target,k)
length(knn.m1)

cm<-table(leaf.data.test.target,knn.m1)
cm
```
Transforming the data by scaling or normalizing gives better clustering as we can observe the same with the above examples. With the raw data we can see leaf 1 is clustered in two separate clusters i.e. there seems to be a tie but we can see with the scaled and normalized examples the same is clustered in only 1 cluster.

That is we can say by normalizing or scaling the data we are gaining confidence to cluster data better.