Risky-Business.Rmd

---
title: "Risky Business"
author: "Jack Carter"
date: "2/21/2022"
output: github_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
library(tidyverse)
library(dplyr)
library(jsonlite)
library(knitr)
library(ggplot2)
library(ggthemes)

# reads in the risk headlines. 
data <- read_csv("C:\\Users\\HUAWEI\\Desktop\\Projects\\Risky-Business\\Data\\risk_headlines.csv") %>%
  filter(!Category == "regime_instability")

# inserted here to make the code example under 
# the Manually Check Labels section work. 
scores <- data

# gets the z score for a specific set of data points.  
get_z_score <- function(data) {
  mean <- mean(data)
  sigma <- sd(data)
  z_scores <- list()
  for(i in 1:length(data)) {
    z_scores[i] <- (data[i] - mean) / sigma
  }
  return(unlist(z_scores))
}

# get the z-scores for each category of a single continent. 
get_continent_z_scores <- function(continent) {
  data <- data[data["Continent"]==continent,]
  count <- data %>%
    group_by(Category) %>%
    count()
  z_scores <- tibble(z_score = get_z_score(count$n))
  final_data <- count[,1] %>%
    cbind(z_scores) %>%
    mutate(Continent = continent)
  return(final_data)
}

# gets the z-scores for each category by continent.  
z_scores <- rbind_pages(lapply(unique(data$Continent), get_continent_z_scores))

# my personal plot theme for data visualizations. 
my_theme <- theme_economist_white(gray_bg = FALSE) +
  theme(plot.title = element_text(hjust = 0.5,
                                  vjust = 10,
                                  size = 10,
                                  color = "#474747"),
        plot.margin = unit(c(1.5, 1, 1.5, 1), "cm"),
        axis.text = element_text(size = 9,
                                 color = "gray30"),
        axis.text.x=element_text(vjust = -2.5),
        axis.title.x = element_text(size = 9,
                                    color = "gray30",
                                    vjust = -10),
        axis.title.y = element_text(size = 9,
                                    color = "gray30",
                                    vjust = 10),
        legend.direction = "vertical", 
        legend.position = "right",
        legend.title = element_blank(),
        legend.text = element_text(size = 11,
                                   color = "gray20"),
        legend.margin=margin(1, -15, 1, 0),
        legend.spacing.x = unit(0.25, "cm"),
        legend.key.size = unit(1, "cm"), 
        legend.key.height = unit(0.75, "cm"),
        strip.text = element_text(hjust = 0.5,
                                  vjust = 1,
                                  size = 10,
                                  color = "#474747"),
        panel.spacing = unit(2, "lines"))

```

## **Summary**

This project uses web scraping, regex and a labeling mechanism to
uncover the relative distribution of different violence risks in the
emerging world. The results include 822 risk headlines from 140 news sources. Just as Tom Cruise discovered in Risky Business,
those seeking to venture into the unknown could find things get out of
hand quickly if they’re not careful.

&nbsp;

## Results

The results show a strong risk of military conflict in Europe due to the Ukraine crisis and a high risk of internal conflict, such as organized crime, social unrest and terrorism, in regions with weak states like Africa and Latin America. 

```{r, echo = FALSE, message = FALSE, warning = FALSE, dpi=600}

# total bar plot. 
total_data <- data %>%
  group_by(Category, Continent) %>%
  count()

total_data %>%
  ggplot(aes(x=Continent, y=n, fill=Category)) +
  geom_bar(stat = "identity", position = "fill") +
  ggtitle("Continent") +
  ylab("%") +
  xlab("") + 
  my_theme +
  theme(axis.text.y=element_blank()) +
  scale_fill_discrete(name = "Risks", labels = c("Military Conflict", 
                                                 "Organized Crime", 
                                                 "Social Unrest",
                                                 "Terrorism")) +
    scale_x_discrete(guide = guide_axis(n.dodge=2))

```

&nbsp;

## **Disclaimer**

The data above show only the relative distribution of headlines for each region, not differences in the total number of headlines. This means we can only hope to better understand variations in the types of risks in such regions, not how dangerous they are overall.

&nbsp;

## **Method**

### **1) Assemble Links Database:**

A database containing over 250 online newspapers in the emerging world is
assembled.

```{r, echo = FALSE, message = FALSE, warning = FALSE, dpi=600}

# reads in the relevant links data frame. 
read_links_df <- function(region) {
  links <- read_csv(
    paste0(
      "C:\\Users\\HUAWEI\\Desktop\\Projects\\Risky-Business\\data\\Links\\", 
      region, 
      "_links.csv"
    )
  )
  return(links)
}

# reads in the links data frame. 
links <- read_links_df("africa") %>%
  rbind(read_links_df("asia")) %>%
  rbind(read_links_df("europe")) %>%
  rbind(read_links_df("latin_america")) %>%
  rbind(read_links_df("mena"))

# creates a summary table of the total hits for each term. 
summary_table <- links %>%
  group_by(Continent) %>%
  summarise("Total Links"=length(Site)) %>%
  spread(Continent, "Total Links")
kable(summary_table)

```

&nbsp;

### **2) Scrape Headlines:**

The links in the database is scraped several times a week from 17 January 2022 to present.
  
—EXAMPLE CODE SNIPET—

```{r, echo = TRUE, message = FALSE, warning = FALSE, dpi=600}

# reads in the link html text. 
read_link <- function(link) {
  elements <- c("h1", "h2", "h3", "h4", "h5")
  headlines <- read_html(link)
  html <- lapply(elements, html_elements, x = headlines)
  text <- unlist(lapply(html, html_text))
  on.exit(closeAllConnections())
  return(text)
}

```

&nbsp;

### **3) Regex:**

A series of regex strings, including 1) keywords, 2) actors and actions, 3) and hyperbolic situations, are combined to calculate the scores of four different kinds of violence.
  
—EXAMPLE CODE SNIPET—

```{r, echo = TRUE, message = FALSE, warning = FALSE, dpi=600}

# strong capture words.
strong_capture_words <-  paste0(
  "(",
  capture <-  "captur(e|es|ed|ing)\\b|",
  hijack <-  "hijac(k|ks|ked|king)\\b|",
  hold_prisoner <-  "h(old|olds|eld|olding) (hostag(e|es)|to ransom)\\b|",
  take_prisoner <-  "t(ake|akes|ook|aken|aking) (prisoner|hostage)\\b|",
  abduct <-  "abduc(t|ts|ted|ting)\\b|",
  kidnap <-  "kidna(p|ps|ped|ping)\\b",
  ")"
)

```

&nbsp;

### **4) Labeling:**

The headlines are assigned one of the four violence type labels, including military conflict, organized crime, social unrest and terrorism. 

—EXAMPLE CODE SNIPET—

```{r, echo = TRUE, message = FALSE, warning = FALSE, dpi=600}

# determines the main category under which the headlines fall. 
assign_category <- function(scores) {
  data <- tibble()
  for(i in 1:nrow(scores)) {
    data[i,1] <- if(sum(scores[i,1:5])==0) "None"
    else {
      categories[max(which(scores[i,1:5]==max(scores[i,1:5])))]
    } 
  }
  return(data)                
}

```

&nbsp;

### **5) Check & Re-assign:**

The top 100 scoring headlines are checked each time the data is scraped, with only those deemed genuine violence risks kept and those mislabeled re-assigned. 

—EXAMPLE CODE SNIPET—

```{r, echo = TRUE, message = FALSE, warning = FALSE, dpi=600}

# creates a list of mislabeled headlines. 
mislabeled <- tibble(
  number = c(2),
  category = c("organized_crime")
)

# adjusts mislabeled headlines. 
mislabeled_adjust <- function(scores) {
  scores <- scores
  for(i in 1:nrow(mislabeled))
    scores[pull(mislabeled[i,1]),3] <- mislabeled$category[i]
  return(scores)
}
mislabeled_adjusted <- mislabeled_adjust(scores)

```


&nbsp;

## **Sources**

- Chong (2021) https://towardsdatascience.com/regular-expressions-clearly-explained-with-examples-822d76b037b4

- The Economist (2021) https://www.economist.com/leaders/2021/07/31/unrest-and-economic-underperformance-stalk-the-emerging-world

- UN (2020) https://www.un.org/en/un75/new-era-conflict-and-violence

- World Bank (2022) https://www.worldbank.org/en/topic/fragilityconflictviolence/overview#1