Lecture3_MachineLearning_Unsupervised.Rmd

title: "Lecture 3 - Unsupervised_Learning"
output: html_document
date: "2023-09-27"
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```

#Part 2 - Unsupervised analysis

##Install packages as needed
```{r}
packages_to_install=c("ggplot2","ggrepel","dendextend")
install.packages(setdiff(packages_to_install, rownames(installed.packages())))
rm(packages_to_install)
```

##Load packages
```{r}
require(ggplot2)
require(ggrepel)
require(dendextend)
```

##Load animal attribute data set
```{r}
animaldata=read.csv("Lecture3_animaltable.csv",as.is=T,header=T,row.names=1)
```

##Explore the data set using "head" or "summary" or in the workspace
```{r}
##Write your code here
head(animaldata)
table(animaldata$type)
```

##Step 1: separate categories from data matrix, assign colors to each category
```{r}
animalmat=animaldata[,-17]
animalcategories=animaldata[,17]
names(animalcategories)=rownames(animalmat)
colvec=c("red","orange","darkgreen","blue","steelblue","purple","black")
plotcolors=colvec[as.numeric(as.factor(animalcategories))]
```

##Step 2: Dimensionality reduction using PCA
```{r}
pr1=prcomp(animalmat)
```

##Explore the PCA output using head/summary/environment window
```{r}
##Write your code here
```

##Step 3: Examine the proportion of variance
```{r}
pcsummary=summary(pr1)
barplot(pcsummary$importance[2,])
```

##Step 4: Examine feature loadings for PC1
```{r}
barplot(pr1$rotation[,1],las=2,ylab="PC1 loading")
```

##Exercise: Examine feature loadings for PC2, PC3, and PC4
```{r}
##Write your code here
```


##Step 4: Plot the animals in PC space
```{r}
###This is complicated plotting code using ggplot - it allows for much more flexibility, but takes some practice to get used to
pcs_to_plot=c(1,2)

dat = data.frame(pr1$x[,pcs_to_plot])
dat$animal=rownames(dat)
p <- ggplot(dat, aes(PC1, PC2, label = animal)) +
  geom_point(color = plotcolors)
p2 <- p + geom_text_repel(max.overlaps = 80) + theme_bw()
p2
```

##Exercise: try plotting PC2 and PC3, and then PC3 and PC4
```{r}
###Modify the code below
pcs_to_plot=c(1,2)

dat = data.frame(pr1$x[,pcs_to_plot])
dat$animal=rownames(dat)
p <- ggplot(dat, aes(PC1, PC2, label = animal)) +   ##this line also needs to be modified
  geom_point(color = plotcolors)
p2 <- p + geom_text_repel(max.overlaps = 80) + theme_bw()
p2
```


##Step 5: Testing out clustering of animals using hierarchical clustering
```{r}
##Hierarchical clustering without dimensionality reduction
###Specify a distance metric and a grouping method - here, use Euclidean distance and complete linkage
#By default, the hcluster function clusters the rows of a data.frame
distance_values = dist(animalmat,method = "euclidean")
clustering = hclust(distance_values,method = "complete")
dendrogram1 = as.dendrogram(clustering)
labels_colors(dendrogram1) = plotcolors[order.dendrogram(dendrogram1)]
plot(dendrogram1)
```

###Try playing around with various grouping methods
```{r}
grouping_method = "single"   ###possible choices: "single", "average", "complete", "ward.D"
distance_values = dist(animalmat,method = "euclidean")
clustering = hclust(distance_values,method = grouping_method)
dendrogram1 = as.dendrogram(clustering)
labels_colors(dendrogram1) = plotcolors[order.dendrogram(dendrogram1)]
plot(dendrogram1)
```

##Cutting the tree to generate a specific number of clusters
```{r}
###Generate hierarchical tree
distance_values = dist(animalmat,method = "euclidean")
clustering <- hclust(distance_values,method = "complete")
dendrogram1 = as.dendrogram(clustering)
labels_colors(dendrogram1) = plotcolors[order.dendrogram(dendrogram1)]
plot(dendrogram1)


###Cut the tree to get 7 clusters
clusters <- cutree(clustering,k = 7)
hclust_correlation_complete_7=data.frame(cluster=clusters[clustering$order])
hclust_correlation_complete_7
```

##Cutting the tree at a specific height (distance metric)
```{r}
###Generate hierarchical tree
distance_values = dist(animalmat,method = "euclidean")
clustering <- hclust(distance_values,method = "complete")
dendrogram1 = as.dendrogram(clustering)
labels_colors(dendrogram1) = plotcolors[order.dendrogram(dendrogram1)]
plot(dendrogram1)

###Cut the tree at a specific height (3)
abline(h=3,lty=2)
clusters <- cutree(clustering, h=3)
hclust_correlation_complete_3=data.frame(cluster=clusters[clustering$order])
hclust_correlation_complete_3
```

##Check how well these clusters recapitulate the known classes
```{r}
table(hclust_correlation_complete_7[,1],animalcategories[rownames(hclust_correlation_complete_7)])
```




##K-means clustering
###The default distance metric is the Euclidean distance, whereas the algorithm can be selected
###Note that k-means clustering has randomness, since the initial cluster centers are selected randomly - to ensure reproducibility, need to set a random seed
```{r}
set.seed(0) ###setting the random seed
###k-means clustering, specifying 7 clusters
clustering_kmeans=kmeans(animalmat,center=7)
kmeans_7=data.frame(cluster=clustering_kmeans$cluster[order(clustering_kmeans$cluster)])
kmeans_7
```
##Check how well these clusters recapitulate the known classes
```{r}
table(kmeans_7[,1],animalcategories[rownames(kmeans_7)])
```
##Visualize k means clusters on the reduced dimension PCA plot
```{r}
pcs_to_plot=c(1,2)

dat = data.frame(pr1$x[,pcs_to_plot])
dat$animal=rownames(dat)
p <- ggplot(dat, aes(PC1, PC2, label = animal)) +   
  geom_point(color = plotcolors)
p2 <- p + geom_text_repel(max.overlaps = 80) + theme_bw() + ggtitle("Original classes")
p2


kmeans_colors=colvec[kmeans_7[rownames(dat),1]]
  
p <- ggplot(dat, aes(PC1, PC2, label = animal)) +   ##this line also needs to be modified
  geom_point(color = kmeans_colors)
p2 <- p + geom_text_repel(max.overlaps = 80) + theme_bw() + ggtitle("k-means clusters")
p2



```