STAT154_rmd.Rmd

---
output:
  pdf_document: default
  html_document: default
---

1 | Data Description
------------

We will analyze the data on fraudulent behavior of job post. The dataset contains information for 5,362 observations that
defines job posting. These job postings are categorized as either fake or real. The dataset is highly unbalanced, with 259(4.8% of the jobs) being
fraudulent jobs and 5,103(around 95.2% of the jobs) of the real jobs. For each job posting recorded, the dataset contains the following 18 
variables where *fraudulent* is treated as our target variable:

|  Variable Name | Type | Description   |
--------|------|---------------|
| `job_id` | int | Unique Job ID |
| `title` | text | The title of the job ad entry |
| `location` | text | Geographical location of the job ad |
| `department` | text | Corporate department |
| `salary_range` | text | Expected salary range |
| `company_profile` | text | Company description |
| `description` | text | The details description of the job ad |
| `requirements` | text | Enlisted requirements for the job opening |
| `benefits` | text | Enlisted offered benefits by the employer |
| `telecommuting` | boolean | output class [1: telecommuting, 0: Not] |
| `has_company_logo` | boolean | output class [1: has company logo, 0: Not] |
| `has_questions` | boolean | output class [1: has questions, 0: Not] |
| `employment_type` | text | Employment type e.g(Full-time, Part-time) |
| `required_experience` | text | Required experience e.g(Intership) |
| `required_education` | text | Required education level |
| `industry` | text | Job industry |
| `function` | text | job position |
| `fraudulent` | boolean | output class [1: fraudulent, 0: Not] |

```{r setup, include=FALSE, message = FALSE}
library(tidyverse)
library(ggplot2)
library(ggplot2)
library(gridExtra)
library(dplyr)
library(plyr)
job <- read_csv("~/Desktop/job_training_data.csv")
```

```{r echo=FALSE, fig.height=3, message = FALSE}
#bivariate distributions
#factor
job$telecommuting = as.factor(job$telecommuting)
job$has_company_logo = as.factor(job$has_company_logo)
job$has_questions = as.factor(job$has_questions)
job$fraudulent = as.factor(job$fraudulent)

#filter location
#telecommuniting 
bar_tele = ggplot(job, aes(x = telecommuting, 
           fill = fraudulent)) + geom_bar(position = "dodge")

#has_company_logo
bar_logo = ggplot(job, aes(x = has_company_logo, 
           fill = fraudulent)) + geom_bar(position = "dodge")

#has_questions
bar_questions = ggplot(job, aes(x = has_questions, 
           fill = fraudulent)) + geom_bar(position = "dodge") 

grid.arrange(bar_tele, bar_logo, bar_questions, 
             ncol=3, nrow=1)

```

The graph above shows most jobs do not require telecommuting, and has company logo. *has_question* graphs shows almost same value for has question 
or not if not fraudulent.

```{r fig.cap="Distributions", echo=FALSE, fig.height=4, message=FALSE}
library(ggplot2)
library(dplyr)
library(tidyverse)
library(ggthemes)
library(skimr)
library(DataExplorer)
library(ggpubr)
#employment type

fjobemp = job %>% filter(fraudulent == 1) %>% group_by(employment_type) %>% dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>%
  ggplot(aes(x = reorder(employment_type, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Fraud Job - Employment type") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

nfjobemp = job %>% filter(fraudulent == 0) %>% group_by(employment_type) %>% 
  dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>%
  ggplot(aes(x = reorder(employment_type, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Non Fraud Job - Employment type") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

ggarrange(fjobemp, nfjobemp)
#required experience

fjobexp = job %>% filter(fraudulent == 1) %>% group_by(required_experience) %>% dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(required_experience, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Fraud Job - Required Experience") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

nfjobexp = job %>% filter(fraudulent == 0) %>% group_by(required_experience) %>% 
  dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(required_experience, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Non Fraud Job - Required Experience") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

ggarrange(fjobexp, nfjobexp)
#Fraud job required education

fjobedu <- job %>% filter(fraudulent == 1) %>% group_by(required_education) %>% dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% 
  ggplot(aes(x = reorder(required_education, -Freq), y = Freq)) + geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Fraud Job - Required Education",
                    x = "Education",
                    
                    y = "Count") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

nfjobedu <- job %>% filter(fraudulent == 0) %>% group_by(required_education)%>% dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% 
  ggplot(aes(x = reorder(required_education, -Freq), y = Freq)) + geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Non Fraud Job - Required Education",
                    x = "Education",
                    y = "Count") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

ggarrange(fjobedu, nfjobedu)
#industry
fjobind = job %>% filter(fraudulent == 1) %>% group_by(industry) %>% dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(industry, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Fraud Job - Industry",
                    x = "Industry", y = "Count") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

nfjobind = job %>% filter(fraudulent == 0) %>% group_by(industry) %>% 
  dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(industry, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Non Fraud Job - Industry",
                    x = "Industry", y = "Count") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 
ggarrange(fjobind, nfjobind)
#fuction

fjobfunc = job %>% filter(fraudulent == 1) %>% group_by(function.) %>% dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(function., -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Fraud Job - Function.") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

nfjobfunc = job %>% filter(fraudulent == 0) %>% group_by(function.) %>% 
  dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(function., -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Non Fraud Job - function.") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

ggarrange(fjobfunc, nfjobfunc)
#department
fjobdep = job %>% filter(fraudulent == 1) %>% group_by(department) %>% dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(department, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Fraud Job - Department") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 

nfjobdep = job %>% filter(fraudulent == 0) %>% group_by(department) %>% 
  dplyr::summarize(Freq = n()) %>% arrange(desc(Freq)) %>% slice(2:11) %>%
  ggplot(aes(x = reorder(department, -Freq), y = Freq)) + 
  geom_bar(stat = "identity", color = "black", fill = "pink") + 
  theme_bw() + labs(title = "Non Fraud Job - Department") +
  geom_text(aes(label=round(Freq,0)), vjust= -0.2) + 
  theme(axis.text.x=element_text(size=10, angle=90,hjust=0.5,vjust=1)) 
ggarrange(fjobdep, nfjobdep)

```

2 | Feature Creation
------------

I created the corpus objects for each complex text feature(title, company_profile, description, requirements, benefits). To clean the text features, I converted all words to lowercase and then removed numbers, punctuations, stop words and stripped extra whitespaces. After that process, I conducted stemming to extract stems for the words and made the term frequency matrix. I removed sparse terms by setting the maximal allowed sparsity to 0.8. This is because when I set higher values as the maximal allowed sparsity, there were so many features that made the computation so expensive. I also indicated which column of the original dataset the word feature came from in the column name of each word feature. I combined the word features with the binary features and the dummy variables from categorical features. As a result, the dimension of my feature matrix is (5362, 82). This means I end up with 82 features as of now. 


```{r, echo=FALSE, results='hide', message = FALSE, warning=FALSE}
job <- read_csv("~/Desktop/job_training_data.csv")
fraud <- job %>% filter(fraudulent == 1)
not_fraud <- job %>% filter(fraudulent == 0)
head(fraud)
head(not_fraud)
dim(fraud)
dim(not_fraud)
colSums(is.na(job))
```


```{r, echo=FALSE, results='hide', message = FALSE, warning=FALSE}
# Create the dummy variables for categorical data(employment_type, required_experience, required_education)
library(fastDummies)
categorical <- subset(job, select = c(employment_type, required_experience, required_education))
ctgr_dummy <- dummy_cols(categorical, select_columns=c('employment_type', 'required_experience', 'required_education'),  remove_most_frequent_dummy = TRUE, remove_selected_columns = TRUE)
ctgr_dummy[is.na(ctgr_dummy)] <- 0
colnames(ctgr_dummy) <- gsub(" ", ".", colnames(ctgr_dummy))
colnames(ctgr_dummy) <- gsub("-", ".", colnames(ctgr_dummy))
colnames(ctgr_dummy) <- gsub("'", ".", colnames(ctgr_dummy))
```

```{r, echo=FALSE, results='hide', message = FALSE, warning=FALSE}
# Parse out the word features from the complex text features(title, company_profile, description, requirements, benefits)

library(tm)
library(SnowballC)
library(tidytext)

title_corpus <- VCorpus(VectorSource(job$title))
profile_corpus <- VCorpus(VectorSource(job$company_profile))
description_corpus <- VCorpus(VectorSource(job$description))
requirements_corpus <- VCorpus(VectorSource(job$requirements))
benefits_corpus <- VCorpus(VectorSource(job$benefits))

clean_text <- function(corpus){
  corpus <- tm_map(corpus, content_transformer(tolower), lazy = T)
  corpus <- tm_map(corpus, removeNumbers, lazy = T)
  corpus <- tm_map(corpus, removePunctuation, lazy = T)
  corpus <- tm_map(corpus, removeWords, stopwords(kind = "en"), lazy = T )
  corpus <- tm_map(corpus, stripWhitespace, lazy = T )
  corpus <- tm_map(corpus, stemDocument, lazy = T)
  corpus <- tm_map(corpus, stripWhitespace, lazy = T)
  word_freq <- DocumentTermMatrix(corpus)
  remove_sparse <- removeSparseTerms(word_freq, 0.8)
  remove_sparse_df <- as.data.frame(as.matrix(remove_sparse))
  return(remove_sparse_df)
}


cleaned_title <- clean_text(title_corpus)
if(ncol(cleaned_title) > 0){
  cleaned_title <- cleaned_title %>% rename_all(paste0, "_title")
}


cleaned_profile <- clean_text(profile_corpus)
if(ncol(cleaned_profile) > 0){
  cleaned_profile <- cleaned_profile %>% rename_all(paste0, "_profile")
}

cleaned_description <- clean_text(description_corpus)
if(ncol(cleaned_description) > 0){
  cleaned_description <- cleaned_description %>% rename_all(paste0, "_description")
}

cleaned_requirements <- clean_text(requirements_corpus)
if(ncol(cleaned_requirements) > 0){
  cleaned_requirements <- cleaned_requirements %>% rename_all(paste0, "_requirements")
}

cleaned_benefits <- clean_text(benefits_corpus)
if(ncol(cleaned_benefits) > 0){
  cleaned_benefits <- cleaned_benefits %>% rename_all(paste0, "_benefits")
}

```

```{r, echo=FALSE, results='hide', message = FALSE, warning=FALSE}
# Combine features
# the word features (from title, company_profile, description, requirements, benefits) 
# + binary features (telecommuting, has_company_logo, has_questions) 
# + categorical features (employment_type, required_experience, required_education)

library(plyr)
cleaned_description$fraudulent <- job$fraudulent 
features <- cbind(cleaned_title, cleaned_profile, cleaned_description, cleaned_requirements, cleaned_benefits, subset(job, select = c(telecommuting, has_company_logo, has_questions)), ctgr_dummy)
```

3 | Unsupervised Feature Filtering
------------


```{r, echo=FALSE, results='hide', fig.cap="Stepwise Selection", fig.height=3, message = FALSE, warning=FALSE}
# Feature selection using stepwise selection

features_model = lm(fraudulent ~., data = features)
library(MASS)
library(olsrr)
ols_model = ols_step_both_aic(features_model, details= TRUE)
ols_model$predictors
plot(ols_model)
```

In order to filter the features provided from the above process, we used the stepwise selection process and selected the best model using cross-validated prediction error of AIC. There was no problem using the “ols_step_both_aic” function from the olsrr package with our feature dimension, and it yielded a better result than the forward and backward selection. By applying this method, we reduced are features to 43 making our matrix dimension (5362, 43).

4 | Power Feature Creation
------------

```{r echo=FALSE, message = FALSE, warning=FALSE}

library(tidyverse)

make_power_features <- function(df) {
  
  # country, state
  split_location <- strsplit(df$location, ", ")
  
  country <- c()
  for (i in 1:length(split_location)) {
    country <- c(country, split_location[[i]][1])
  }
  country[is.na(country)] <- 0
  
  state <- c()
  for (i in 1:length(split_location)) {
    state <- c(state, split_location[[i]][2])
  }
  state[is.na(state)] <- 0
  
  country_state <- data.frame(country = country, state = state)
  country_state[country_state$country != "US", "state"] <- ""
  
  # state_TX
  state_TX <- country_state$state
  state_TX <- ifelse(state_TX == "TX", 1, 0)
  #state_TX[state_TX != "TX"] <- 0
  #state_TX[state_TX == "TX"] <- 1
  state_TX = as.numeric(state_TX)
  
  # state_NY
  state_NY <- country_state$state
  state_NY <- ifelse(state_NY == "NY", 1, 0)
  #state_NY[state_NY != "NY"] <- 0
  #state_NY[state_NY == "NY"] <- 1
  state_NY = as.numeric(state_NY)
  
  # state_CA
  state_CA <- country_state$state
  state_CA <- ifelse(state_CA == "CA", 1, 0)
  #state_CA[state_CA != "CA"] <- 0
  #state_CA[state_CA == "CA"] <- 1
  state_CA = as.numeric(state_CA)
  
  # length_des
  length_des <- nchar(df$description)
  length_des[is.na(length_des)] <- 0
  
  # length_ben
  length_ben <- nchar(df$benefits)
  length_ben[is.na(length_ben)] <- 0
  
  # contain_email
  contain_email <- str_extract(df$company_profile, "#PHONE_(.*)#")
  contain_email[is.na(contain_email)] <- 0
  contain_email[contain_email != "0"] <- 1
  contain_email = as.numeric(contain_email)
  
  # length_req
  length_req <- nchar(df$requirements)
  length_req[is.na(length_req)] <- 0
  
  # contain_phone
  contain_phone <- str_extract(df$company_profile, "#EMAIL_(.*)#")
  contain_phone[is.na(contain_phone)] <- 0
  contain_phone[contain_phone != "0"] <- 1
  contain_phone = as.numeric(contain_phone)
  
  # has_salary
  has_salary <- ifelse(is.na(df$salary_range), 0, 1)
  
  oil_ind <- ifelse(df$industry == "Oil & Energy", 1, 0)
  oil_ind[is.na(oil_ind)] <- 0

  hos_ind <- ifelse(df$industry == "Hospital & Health Care", 1, 0)
  hos_ind[is.na(hos_ind)] <- 0

  acc_ind <- ifelse(df$industry == "Accounting", 1, 0)
  acc_ind[is.na(acc_ind)] <- 0

  oil_dept <- ifelse(df$department == "Oil & Energy", 1, 0)
  oil_dept[is.na(oil_dept)] <- 0
  
  length_profile <- nchar(df$company_profile)
  length_profile[is.na(length_profile)] <- 0

  eng_dept <- ifelse(df$department == "Engineering", 1, 0)
  eng_dept[is.na(eng_dept)] <- 0
  
  #customer_dept <- ifelse(df$department == "Customer Service", 1, 0)
  #customer_dept[is.na(customer_dept)] <- 0
  
  #clerical_dept <- ifelse(df$department =="Clerical", 1, 0)
  #clerical_dept[is.na(clerical_dept)] <- 0
  
  #acc_dept <- ifelse(df$department == "Account", 1, 0)
  #acc_dept[is.na(acc_dept)] <- 0
  
  #admin_dept <- ifelse(df$department == "admin", 1, 0)
  #admin_dept[is.na(admin_dept)] <- 0
  
  # uppercase_des
  upper_des <- lengths(str_extract_all(df$description, "[A-Z]{3,}+"))
  upper_des[is.na(upper_des)] <- 0


# uppercase_req
  upper_req <- lengths(str_extract_all(df$requirements, "[A-Z]{3,}+"))
  upper_req[is.na(upper_req)] <- 0


# uppercase_ben
  upper_ben <- lengths(str_extract_all(df$benefits, "[A-Z]{3,}+"))
  upper_ben[is.na(upper_ben)] <- 0
  
# star_des
  star_des <- as.numeric(grepl("*", df$description, fixed = TRUE))
  star_des[is.na(star_des)] <- 0


# star_req
  star_req <- as.numeric(grepl("*", df$requirements, fixed = TRUE))
  star_req[is.na(star_req)] <- 0

# star_ben
  star_ben <- as.numeric(grepl("*", df$benefits, fixed = TRUE))
  star_ben[is.na(star_ben)] <- 0

# na_company
  na_company <- as.numeric(is.na(df$company_profile))

  
  features <- data.frame(state_TX = state_TX,
                         length_des = length_des,
                         length_ben = length_ben,
                         state_NY = state_NY,
                         state_CA = state_CA,
                         contain_email = contain_email,
                         length_req = length_req,
                         contain_phone = contain_phone,
                         has_salary = has_salary,
                         oil_ind = oil_ind,
                         hos_ind = hos_ind,
                         acc_ind = acc_ind,
                         oil_dept = oil_dept,
                         length_profile = length_profile,
                         eng_dept = eng_dept,
                         upper_des = upper_des,
                         upper_req = upper_req,
                         upper_ben = upper_ben
                         )
  
  return(features)
  
}
```




```{r}
# power features
power_features <- make_power_features(job)
```

5 | Feature and Power Feature Combination
------------
```{r}
final_features = cbind(subset(features, select = ols_model$predictors), power_features)
```



```{r}
# Split the dataset into train and test
final_features$fraudulent <- job$fraudulent

set.seed(12345)

split <- sample(c(TRUE, FALSE), nrow(final_features), replace=TRUE, prob=c(0.8, 0.2))
job_train <-  final_features[split,]
job_test <- final_features[!split, ]

job_train
job_test


job_train$fraudulent = as.factor(job_train$fraudulent)
job_test$fraudulent = as.factor(job_test$fraudulent)

final_features$fraudulent <- as.factor(job$fraudulent)
```

```{r, echo=FALSE, results='hide'}
dim(final_features)
write.csv(final_features, "final_features.csv")
```
8 | Clustering
------------

```{r, echo=FALSE, results='hide', message=FALSE, warning='hide'}
library(tidyverse)
library(cluster)
library(haven)
library(ggdendro)
library(NbClust)
library(factoextra)
library(klaR)
library(data.table)
library(rlang)
library(dplyr)
library(NbClust)
library(ggpubr)
library(corrplot)
ff = subset(final_features, select = -fraudulent) %>% scale()

ncol <- ncol(ff)
# Store variable names
var <- list()
for(i in 1:ncol){
   var[[i]] <- names(ff)[i]
}
# Rename columns for easier use
names(ff)[1:ncol] <- paste("var", 1:ncol, sep="")


######clustering
### Hierarchical Clustering 
## Determine the number of clusters(after scale)

# Elbow method
fviz_nbclust(ff, kmeans, method = "wss") +
    geom_vline(xintercept = 4, linetype = 2)+
  labs(subtitle = "Elbow method")

# Silhouette method
fviz_nbclust(ff, kmeans, method = "silhouette")+
  labs(subtitle = "Silhouette method")


# set.seed(123)
#final = kmeans(ff, 2, nstart =25)
#print(final)
#fviz_cluster(final, data = ff)

#final = kmeans(ff, 3, nstart =25)
#print(final)
#fviz_cluster(final, data = ff)

#final = kmeans(ff, 4, nstart =25)
#print(final)
#fviz_cluster(final, data = ff)

#kmeans = as.data.frame(final$cluster)
#colnames(kmeans)[1] = "Cluster"
#kmeans = cbind(kmeans, fraudulent = job$fraudulent)

#nrow(filter(kmeans, Cluster == 1))
#nrow(filter(kmeans, Cluster == 2))
#nrow(filter(kmeans, Cluster == 3))
#nrow(filter(kmeans, Cluster == 4))

#sum(filter(kmeans, Cluster == 1)$fraudulent)
#sum(filter(kmeans, Cluster == 1)$fraudulent)/sum(filter(job, fraudulent == 1)$fraudulent)
#sum(filter(kmeans, Cluster == 1)$fraudulent)/nrow(filter(databind, Cluster == 1))

#sum(filter(kmeans, Cluster == 2)$fraudulent)
#sum(filter(kmeans, Cluster == 2)$fraudulent)/sum(filter(job, fraudulent == 1)$fraudulent)
#sum(filter(kmeans, Cluster == 2)$fraudulent)/nrow(filter(databind, Cluster == 2))

#sum(filter(kmeans, Cluster == 3)$fraudulent)
#sum(filter(kmeans, Cluster == 3)$fraudulent)/sum(filter(job, fraudulent == 1)$fraudulent)
#sum(filter(kmeans, Cluster == 3)$fraudulent)/nrow(filter(databind, Cluster == 3))

#sum(filter(kmeans, Cluster == 4)$fraudulent)
#sum(filter(kmeans, Cluster == 4)$fraudulent)/sum(filter(job, fraudulent == 1)$fraudulent)
#sum(filter(kmeans, Cluster == 4)$fraudulent)/nrow(filter(databind, Cluster == 4))
```

9 | Improve Features
------------


These are power features for improving the model. We selected some of these features that showed significant difference to classify fraudulent job post.

```{r, echo=FALSE, results='hide', message=FALSE, warning='hide'}
fraud <- features[features$fraudulent == 1, ]
real <- features[features$fraudulent == 0, ]

TX_prop_fraud <- sum(fraud$state_TX) / dim(fraud)[1]
TX_prop_real <- sum(real$state_TX) / dim(real)[1]

CA_prop_fraud <- sum(fraud$state_CA) / dim(fraud)[1]
CA_prop_real <- sum(real$state_CA) / dim(real)[1]

NY_prop_fraud <- sum(fraud$state_NY) / dim(fraud)[1]
NY_prop_real <- sum(real$state_NY) / dim(real)[1]

state_mat <- matrix(c(TX_prop_fraud, TX_prop_real, CA_prop_fraud, CA_prop_real, NY_prop_fraud, NY_prop_real), nrow = 2)
state_df <- as.data.frame(state_mat)
colnames(state_df) <- c("TX", "CA", "NY")
rownames(state_df) <- c("fraud", "real")
state_mat <- as.matrix(state_df)

barplot(state_mat,
        beside = TRUE,
        main = "State: Fraud vs Real",
        col = c("#F39B7FFF", rgb(0.2, 0.4, 0.6, 0.6)),
        ylim = c(0, 0.3))
legend("topright",
       c("Fraud", "Real"),
       fill = c("#F39B7FFF", rgb(0.2, 0.4, 0.6, 0.6)))
```

There is a difference in proportion of state between fraudulent job post and real job post. Many job posts were from Texas and the proportion was greater than the real job post.

```{r, echo=FALSE, results='hide', message=FALSE, warning='hide'}
boxplot(features$length_req ~ features$fraudulent,
        ylim = c(0, 2000),
        main = "Length of Requirement: Fraud vs Real",
        names = c("real", "fraud"),
        ylab = "length of requirement",
        xlab = "",
        col = c("khaki", "honeydew4"))
```

The length of requirement was longer in the real job post. 

```{r, echo=FALSE, results='hide', message=FALSE, warning='hide'}
oil_ind_fraud <- sum(fraud$oil_ind) / dim(fraud)[1]
oil_ind_real <- sum(real$oil_ind) / dim(real)[1]

hos_ind_fraud <- sum(fraud$hos_ind) / dim(fraud)[1]
hos_ind_real <- sum(real$hos_ind) / dim(real)[1]

acc_ind_fraud <- sum(fraud$acc_ind) / dim(fraud)[1]
acc_ind_real <- sum(real$acc_ind) / dim(real)[1]

ind_mat <- matrix(c(oil_ind_fraud, oil_ind_real, hos_ind_fraud, hos_ind_real, acc_ind_fraud, acc_ind_real), nrow = 2)
ind_df <- as.data.frame(ind_mat)
colnames(ind_df) <- c("Oil & Energy", "Hospitality", "Accounting")
rownames(ind_df) <- c("fraud", "real")
ind_mat <- as.matrix(ind_df)

barplot(ind_mat,
        beside = TRUE,
        main = "Industry: Fraud vs Real",
        col = c("palegreen3", "dodgerblue3"),
        ylim = c(0, 0.15))
legend("topright",
       c("Fraud", "Real"),
       fill = c("palegreen3", "dodgerblue3"))

```

There is a significant difference in proportion of industry between fraudulent job post and real job post.


10 | Validation Set
------------

```{r, echo=FALSE, results='hide', message=FALSE, warning='hide'}
# SVM
library(e1071)
svm_model <- svm(fraudulent~., data=job_train , kernel ="radial", scale=TRUE)
summary(svm_model)

train_pred_svm <- predict(svm_model, subset(job_train, select = -fraudulent))
table(train_pred_svm, job_train$fraudulent)

test_pred_svm <- predict(svm_model, subset(job_test, select = -fraudulent))
table(test_pred_svm, job_test$fraudulent)
```

```{r, echo=FALSE, results='hide', message=FALSE, warning='hide'}
# Random Forest Model
library(randomForest)

rf_model <- randomForest(fraudulent~., data=job_train, ntree=120, mtry = 25, importance =TRUE)

train_pred_rf <- predict(rf_model, subset(job_train, select = -fraudulent))
table(train_pred_rf, job_train$fraudulent)

test_pred_rf <- predict(rf_model ,subset(job_test, select = -fraudulent))
table(test_pred_rf, job_test$fraudulent)

```

```{r, echo=FALSE, results='hide', message=FALSE, warning='hide'}
# Random Forest Model
library(randomForest)

rf_model <- randomForest(fraudulent~., data=final_features, cutoff = c(0.8, 0.2), mtry = 23, importance =TRUE)

saveRDS(rf_model, file = "rf_model1.RDS")
```