# Set the folder where the files are contained
folder <- "C:\\Users\\Brett\\Dropbox\\CUNY\\621\\Week6\\"
# Read in the csv and only use complete cases
jd.train <- read.csv(paste0(folder, "jury-training-data(1).csv"))
jd.train <- jd.train[complete.cases(jd.train),]
jd.public <- read.csv(paste0(folder, "jury-learning-data-public(1).csv"))
jd.public <- jd.public[complete.cases(jd.public),]
jd.private <- read.csv(paste0(folder, "jury-learning-data-private(1).csv"))
jd.private <- jd.private[complete.cases(jd.private),]I rewrote the code, but it is based off previous work.
entropy <- function(d){
# Creates a table of the values and returns each value as its probability
prob <- table(d)/sum(table(d))
# Finds the log base 2 of the probability table and multiplies it by the inverse of the probability then finds that sum
e <- sum((-prob) * log(prob, base=2))
return (e)
}infogain <- function(d, a){
# Calculates the entropy of d
ed <- entropy(d)
# Splits the data frame on d
split.df <- split(d, a)
# Returns the mean value of the entropy function after its been split
s <- mean(sapply(1:length(split.df), function(x) entropy(split.df[x])))
return(ed - s)
}checkgain <- function(input.df, target.var){
# Creates a data frame without the target variable
curr.df <- input.df[-target.var]
# Creates an array of the target variable
curr.var <- input.df[target.var]
# Checks the information gain
temp <- sapply(1:ncol(curr.df), function(x) infogain(curr.df[,x], curr.var))
best.gain.pos <- which.max(temp)
return(as.list(c(best.gain.pos, temp)))
}split.tree <- function(inc.df, var.int){
split.col <- checkgain(inc.df, var.int)[[1]]
split.col.name <- names(inc.df[split.col])
curr.split <- split(inc.df, inc.df[,split.col])
temp <- sapply(1:length(curr.split), function(x) table(curr.split[[x]][,var.int])/nrow(curr.split[[x]]))
colnames(temp) <- names(curr.split)
for (i in 1:length(curr.split)){
curr.split[[i]] <- curr.split[[i]][,-split.col]
}
return(list(split.col.name, temp, curr.split, split.col))
}
first.split <- split.tree(jd.train, 5)build.tree <- function(curr.df, var.int, curr.tree){
if (ncol(curr.df) < 3){
print(c(curr.tree, names(curr.df)[1])[-1])
}
else {
temp.df <- split.tree(curr.df, var.int)
var.int <- var.int - 1
for(i in 1:(length(temp.df[[3]]))){
temp.tree <- c(curr.tree, names(temp.df[[3]])[i])
build.tree(temp.df[[3]][[i]], var.int, temp.tree)
}
}
}
# build.tree(jd.train, 5, array())total.branches <- list()
build.tree <- function(curr.df, var.int, curr.tree){
if (ncol(curr.df) < 3){
temp.df <- split.tree(curr.df, var.int)
var.int <- var.int - 1
for(i in 1:(length(temp.df[[3]]))){
temp.tree <- c(curr.tree, names(temp.df[[3]])[i])
total.branches[[length(total.branches)+1]] <<- temp.tree[-1]
print(temp.tree[-1])
}
}
else {
temp.df <- split.tree(curr.df, var.int)
var.int <- var.int - 1
for(i in 1:(length(temp.df[[3]]))){
temp.tree <- c(curr.tree, names(temp.df[[3]])[i])
build.tree(temp.df[[3]][[i]], var.int, temp.tree)
}
}
}
build.tree(jd.train, 5, array())## [1] "Female" "Divorced" "Older Adult" "Employed"
## [1] "Female" "Divorced" "Older Adult" "Not Employed"
## [1] "Female" "Divorced" "Younger Adult" "Employed"
## [1] "Female" "Divorced" "Younger Adult" "Not Employed"
## [1] "Female" "Married" "Older Adult" "Employed"
## [1] "Female" "Married" "Older Adult" "Not Employed"
## [1] "Female" "Married" "Younger Adult" "Employed"
## [1] "Female" "Married" "Younger Adult" "Not Employed"
## [1] "Female" "Single" "Older Adult" "Employed"
## [1] "Female" "Single" "Older Adult" "Not Employed"
## [1] "Female" "Single" "Younger Adult" "Employed"
## [1] "Female" "Single" "Younger Adult" "Not Employed"
## [1] "Male" "Divorced" "Employed" "Older Adult"
## [1] "Male" "Divorced" "Employed" "Younger Adult"
## [1] "Male" "Divorced" "Not Employed" "Older Adult"
## [1] "Male" "Divorced" "Not Employed" "Younger Adult"
## [1] "Male" "Married" "Employed" "Older Adult"
## [1] "Male" "Married" "Employed" "Younger Adult"
## [1] "Male" "Married" "Not Employed" "Older Adult"
## [1] "Male" "Married" "Not Employed" "Younger Adult"
## [1] "Male" "Single" "Employed" "Older Adult"
## [1] "Male" "Single" "Employed" "Younger Adult"
## [1] "Male" "Single" "Not Employed" "Older Adult"
## [1] "Male" "Single" "Not Employed" "Younger Adult"for(i in 1:length(total.branches)){
if(total.branches[[i]][1] == "Female"){
total.branches[[i]][5] <- names(which.max(table(subset(jd.train,
gender == "Female" &
marital == total.branches[[i]][2] &
agegroup == total.branches[[i]][3] &
employment == total.branches[[i]][4])[,5])))
}
else{
total.branches[[i]][5] <- names(which.max(table(subset(jd.train,
gender == "Male" &
marital == total.branches[[i]][2] &
agegroup == total.branches[[i]][4] &
employment == total.branches[[i]][3])[,5])))
}
}
#total.branchesjd.public[,6] <- "Guilty"
jd.public[jd.public$gender == "Female" &
jd.public$marital == "Married" &
jd.public$agegroup == "Older Adult",6] <- "Not Guilty"
jd.public[jd.public$gender == "Male",6] <- "Not Guilty"
jd.public[jd.public$gender == "Male" &
jd.public$marital == "Married",6] <- "Guilty"
jd.public[jd.public$gender == "Male" &
jd.public$marital == "Married" &
jd.public$agegroup == "Younger Adult" &
jd.public$employment == "Not Employed",6] <- "Not Guilty"
sum(jd.public[,5] == jd.public[,6])/nrow(jd.public)## [1] 0.6806723##### My guesses #####
bretts.classifiers <- jd.private
bretts.classifiers[,5] <- "Guilty"
bretts.classifiers[bretts.classifiers$gender == "Female" &
bretts.classifiers$marital == "Married" &
bretts.classifiers$agegroup == "Older Adult",5] <- "Not Guilty"
bretts.classifiers[bretts.classifiers$gender == "Male",5] <- "Not Guilty"
bretts.classifiers[bretts.classifiers$gender == "Male" &
bretts.classifiers$marital == "Married",5] <- "Guilty"
bretts.classifiers[bretts.classifiers$gender == "Male" &
bretts.classifiers$marital == "Married" &
bretts.classifiers$agegroup == "Younger Adult" &
bretts.classifiers$employment == "Not Employed",5] <- "Not Guilty"
write.csv(bretts.classifiers, paste0(folder, "jd.private.classified.csv"))Nearly 70% accuracy is not fantastic, and it can be tweaked a bit by removing certain branches, but an ideal model isn’t required for the deliverables (and an ideal model would be much easier to create using random forests, etc. anyway)