8. Network analysis

Author

G.H. Koo

Learning goals

By the end of this tutorial, you will be able to:

  • Understand the difference between semantic networks and social networks.
  • Build a semantic network from bigrams in text data.
  • Build a social network from user mentions and replies.
  • Interpret common network measures such as degree, betweenness, and community structure.
  • Visualize networks using igraph and ggraph.

Introduction to Network Analysis

Network analysis helps us study relationships rather than isolated observations. In text analysis, this can mean examining how words co-occur with one another. In social media research, it can mean examining how users interact through replies and mentions.

In this tutorial, we cover two related applications:

  • Semantic network analysis, which maps connections among words
  • Social network analysis, which maps connections among users
Note

A network has two basic parts:

  • nodes, which represent words or users
  • edges, which represent relationships between them

Required Packages

You only need to install packages once.

packages <- c(
  "tidyverse",
  "tidytext",
  "tokenizers",
  "igraph",
  "ggraph",
  "tidygraph",
  "stringr"
)
install.packages(packages)
library(tidyverse)
library(dplyr)
library(tokenizers)
library(tidytext)
library(igraph)
library(ggraph)
library(tidygraph)
library(stringr)
set.seed(381)
options(scipen = 999)

Importing data

For this tutorial, we use a CSV file containing abortion-related Twitter data.

tweets <- read.csv("~/Desktop/abortion_tweets.csv", header = TRUE, sep = ",")

tweets_subset <- tweets %>%
  select(
    id,
    author_id,
    created_at,
    text,
    public_metrics.impression_count,
    public_metrics.like_count,
    public_metrics.quote_count,
    public_metrics.reply_count,
    public_metrics.retweet_count,
    referenced_tweets,
    in_reply_to_user_id,
    entities.mentions
  )

head(tweets_subset)
                   id           author_id               created_at
1 1608879536376262656 1293292839963631616 2022-12-30T17:35:47.000Z
2 1608879534690160640 1407429095005245440 2022-12-30T17:35:47.000Z
3 1608879520039448320  925802396609073280 2022-12-30T17:35:43.000Z
4 1608879519641002240            99833187 2022-12-30T17:35:43.000Z
5 1608879518382718720           228617818 2022-12-30T17:35:43.000Z
6 1608879503165780224          3206943202 2022-12-30T17:35:39.000Z
                                                                                                                                              text
1     RT @KelseyDotOrg: Today my piece for @atrupar’s Public Notice on abortion access in Missouri and Kansas is up. Thank you Aaron for the oppo…
2                                        You know any of these hags? Get Paid!  $10,000 Reward for Info https://t.co/mgpx6PoQTo via @gatewaypundit
3     RT @mjs_DC: After being denied basic miscarriage care during two separate visits to the ER because of Louisiana's abortion ban, Kaitlyn Jos…
4 RT @MrAndyNgo: Jennifer Thompson, an extremist abortion &amp; BLM activist in Portland, has shared in graphic detail her recent decision to end…
5     RT @LifeNewsToo: The FBI has arrested a dozen pro-life Americans for peacefully protesting abortion, but not one single leftist for firebom…
6                                                                              RT @Moneymykpt2: Them Weeks Fly when you tryna get that Abortion 😭
  public_metrics.impression_count public_metrics.like_count
1                               0                         0
2                              10                         1
3                               0                         0
4                               0                         0
5                               0                         0
6                               0                         0
  public_metrics.quote_count public_metrics.reply_count
1                          0                          0
2                          0                          0
3                          0                          0
4                          0                          0
5                          0                          0
6                          0                          0
  public_metrics.retweet_count              referenced_tweets
1                          190 retweeted, 1608817588376834048
2                            1                               
3                         1027 retweeted, 1608871631765786624
4                         2278 retweeted, 1529674574266257408
5                           41 retweeted, 1608860717515423744
6                          367 retweeted, 1603919700890640389
  in_reply_to_user_id
1                  NA
2                  NA
3                  NA
4                  NA
5                  NA
6                  NA
                                                                entities.mentions
1 c(3, 37), c(16, 45), c("KelseyDotOrg", "atrupar"), c("1375165814", "288277167")
2                                                91, 105, gatewaypundit, 19211550
3                                                         3, 10, mjs_DC, 88215673
4                                                    3, 13, MrAndyNgo, 2835451658
5                                                    3, 15, LifeNewsToo, 74552263
6                                         3, 15, Moneymykpt2, 1153357658268819457

Part 1: Semantic Network Analysis

What is a semantic network?

A semantic network shows how words are connected based on their co-occurrence in text. Here, we use bigrams (two-word combinations) to identify words that frequently appear together.

Step 1: Clean retweet prefixes

We begin by removing leading RT @username patterns so they do not dominate the network.

tweets_cleaned <- tweets_subset %>%
  mutate(text = str_remove_all(text, "^RT @\\w+"))

Step 2: Extract bigrams and remove stopwords

tw_bigram <- tweets_cleaned %>%
  unnest_tokens(bigram, text, token = "ngrams", n = 2) %>%
  separate(bigram, into = c("word1", "word2"), sep = " ") %>%
  filter(
    !word1 %in% stop_words$word,
    !word2 %in% stop_words$word
  ) %>%
  unite(bigram, word1, word2, sep = " ") %>%
  count(bigram, sort = TRUE)

head(tw_bigram)
           bigram    n
1      https t.co 2491
2        12 weeks  677
3        pro life  495
4   anti abortion  386
5 abortion rights  286
6    abortion ban  270

Step 3: Separate bigrams into two words

bigram_separated <- tw_bigram %>%
  separate(bigram, into = c("word1", "word2"), sep = " ")

Step 4: Filter infrequent bigrams

bigram_filtered <- bigram_separated %>%
  filter(n > 100)
Tip

The threshold used here (n > 100) is just an example. Lower thresholds keep more word pairs but can make the network harder to read.

Step 5: Build the network object

bigram_graph <- graph_from_data_frame(bigram_filtered)

Step 6: Visualize the semantic network

ggraph(bigram_graph, layout = "fr") +
  geom_edge_link() +
  geom_node_point() +
  geom_node_text(aes(label = name), repel = TRUE) +
  labs(title = "Semantic Network of Bigrams in Abortion Tweets") +
  theme_void()

Optional: Add directional edges

direction_edges <- grid::arrow(type = "closed", length = unit(.2, "cm"))

ggraph(bigram_graph, layout = "fr") +
  geom_edge_link(aes(edge_alpha = n), arrow = direction_edges) +
  geom_node_point(size = 2) +
  geom_node_text(aes(label = name), repel = TRUE) +
  labs(title = "Semantic Network of Bigrams with Directional Information") +
  theme_void()

Describe the semantic network

vcount(bigram_graph)
[1] 42
ecount(bigram_graph)
[1] 32
edge_density(bigram_graph)
[1] 0.01858304
components(bigram_graph)$no
[1] 10

Questions for interpretation

  1. Which words appear most connected?
  2. Are there clusters of words representing different themes?
  3. Do certain words act as bridges between themes?
  4. What does this suggest about how abortion is discussed in the dataset?

Part 2: Social Network Analysis

In social network analysis, users are the nodes and interactions among users are the edges.

Step 1: Prepare interaction data

tweets_interaction <- tweets_subset %>%
  select(author_id, in_reply_to_user_id, entities.mentions)

Step 2: Build reply and mention edge lists

replies <- tweets_interaction %>%
  filter(!is.na(in_reply_to_user_id)) %>%
  transmute(
    source = as.character(author_id),
    target = as.character(in_reply_to_user_id)
  )

mentions <- tweets_interaction %>%
  mutate(
    entities.mentions = as.character(entities.mentions),
    mention_ids = str_extract_all(entities.mentions, "(?<!\\d)\\d{4,}(?!\\d)")
  ) %>%
  select(source = author_id, mention_ids) %>%
  unnest(mention_ids) %>%
  filter(!is.na(mention_ids), mention_ids != "") %>%
  transmute(
    source = as.character(source),
    target = as.character(mention_ids)
  )

edges <- bind_rows(replies, mentions) %>%
  filter(!is.na(source), !is.na(target), source != "", target != "") %>%
  distinct()

head(edges)
               source              target
1          2720585148 1467931973616386048
2          1855645256          1452587928
3 1304024197203550208 1602045892248453120
4           196802928 1428048511547973632
5           171363852            83716493
6 1540777089095401728 1240756589495308288

Step 3: Filter to highly connected users

user_counts <- edges %>%
  pivot_longer(cols = c(source, target), values_to = "user") %>%
  count(user, sort = TRUE)

user_counts %>%
  head(10)
# A tibble: 10 × 2
   user                    n
   <chr>               <int>
 1 88215673              797
 2 1467931973616386052   556
 3 50434933              320
 4 4099171               168
 5 1541493458543775747   162
 6 288277167             134
 7 137472360              97
 8 1014756915094515713    95
 9 1375165814             89
10 1574645533054046209    84
popular_users <- user_counts %>%
  slice_max(order_by = n, n = 30, with_ties = FALSE)

popular_users
# A tibble: 30 × 2
   user                    n
   <chr>               <int>
 1 88215673              797
 2 1467931973616386052   556
 3 50434933              320
 4 4099171               168
 5 1541493458543775747   162
 6 288277167             134
 7 137472360              97
 8 1014756915094515713    95
 9 1375165814             89
10 1574645533054046209    84
# ℹ 20 more rows
Note

In this dataset, a very high threshold can leave too few users to visualize. Here we keep the top 30 most connected users so the network remains readable but still shows more structure.

Step 4: Build and visualize the user network

filtered_edges <- edges %>%
  filter(source %in% popular_users$user, target %in% popular_users$user) %>%
  count(source, target, sort = TRUE, name = "weight")

user_graph <- graph_from_data_frame(filtered_edges, directed = TRUE)

V(user_graph)$degree_total <- degree(user_graph, mode = "all")
V(user_graph)$degree_in <- degree(user_graph, mode = "in")
V(user_graph)$degree_out <- degree(user_graph, mode = "out")

interaction_graph <- ggraph(user_graph, layout = "fr") +
  geom_edge_link(
    aes(width = weight, alpha = weight),
    arrow = arrow(type = "closed", length = unit(0.12, "cm")),
    end_cap = circle(2.5, "mm")
  ) +
  geom_node_point(aes(size = degree_total), color = "steelblue4") +
  geom_node_text(aes(label = name), repel = TRUE, size = 3) +
  scale_edge_width(range = c(0.2, 1.5)) +
  scale_edge_alpha(range = c(0.2, 0.7)) +
  scale_size(range = c(2, 8)) +
  labs(title = "Twitter User Interaction Network") +
  theme_void()

interaction_graph

Interpret user interactions

In this graph:

  • nodes represent Twitter users
  • edges represent mentions or replies
  • more connected nodes indicate users with more ties to others in the network

If you want to inspect a specific user ID, you can sometimes search the platform directly using a format like from:USER_ID, as long as the account is still public and available.

Step 5: Degree and betweenness centrality

centrality_table <- tibble(
  user = names(degree(user_graph, mode = "all")),
  degree_total = as.numeric(degree(user_graph, mode = "all")),
  degree_in = as.numeric(degree(user_graph, mode = "in")),
  degree_out = as.numeric(degree(user_graph, mode = "out")),
  betweenness = as.numeric(betweenness(user_graph, directed = TRUE))
)

centrality_table %>%
  arrange(desc(degree_total)) %>%
  slice_head(n = 10)
# A tibble: 10 × 5
   user                degree_total degree_in degree_out betweenness
   <chr>                      <dbl>     <dbl>      <dbl>       <dbl>
 1 259976840                     13         6          7       6.83 
 2 1311641337175957504           12         6          6       0.833
 3 1589837168842866688           12         9          3      15.5  
 4 33964939                      12         6          6       9    
 5 1293378709093670912           11         6          5       0    
 6 1112147640547008512           10         9          1       0    
 7 1527008268492611840           10         1          9       5.83 
 8 1598856272191062016           10         4          6       0    
 9 988840263056527232             8         1          7       0    
10 1939118838                     4         2          2       1    
centrality_table %>%
  arrange(desc(betweenness)) %>%
  slice_head(n = 10)
# A tibble: 10 × 5
   user                degree_total degree_in degree_out betweenness
   <chr>                      <dbl>     <dbl>      <dbl>       <dbl>
 1 1589837168842866688           12         9          3      15.5  
 2 33964939                      12         6          6       9    
 3 259976840                     13         6          7       6.83 
 4 1527008268492611840           10         1          9       5.83 
 5 1939118838                     4         2          2       1    
 6 1311641337175957504           12         6          6       0.833
 7 1112147640547008512           10         9          1       0    
 8 1202304715280789504            3         0          3       0    
 9 1293378709093670912           11         6          5       0    
10 1598856272191062016           10         4          6       0

Extra: Basic network description

vcount(user_graph)
[1] 17
ecount(user_graph)
[1] 59
edge_density(user_graph)
[1] 0.2169118
components(user_graph)$no
[1] 3

How to interpret these numbers

  • More nodes means more unique users are involved.
  • More edges means more interactions are taking place.
  • Higher density means users are more interconnected.
  • More components means the network contains separate groups that are not connected to each other.

Extra: Compare in-degree and out-degree

in_out_table <- centrality_table %>%
  select(user, degree_in, degree_out) %>%
  arrange(desc(degree_in), desc(degree_out))

top_in_degree <- in_out_table %>%
  arrange(desc(degree_in)) %>%
  slice_head(n = 10)
top_in_degree
# A tibble: 10 × 3
   user                degree_in degree_out
   <chr>                   <dbl>      <dbl>
 1 1589837168842866688         9          3
 2 1112147640547008512         9          1
 3 259976840                   6          7
 4 1311641337175957504         6          6
 5 33964939                    6          6
 6 1293378709093670912         6          5
 7 1598856272191062016         4          6
 8 271432134                   3          0
 9 1939118838                  2          2
10 4099171                     2          1
top_out_degree <- in_out_table %>%
  arrange(desc(degree_out)) %>%
  slice_head(n = 10)
top_out_degree
# A tibble: 10 × 3
   user                degree_in degree_out
   <chr>                   <dbl>      <dbl>
 1 1527008268492611840         1          9
 2 259976840                   6          7
 3 988840263056527232          1          7
 4 1311641337175957504         6          6
 5 33964939                    6          6
 6 1598856272191062016         4          6
 7 1293378709093670912         6          5
 8 1589837168842866688         9          3
 9 1202304715280789504         0          3
10 1939118838                  2          2

How to interpret this

  • High in-degree users are frequently mentioned or replied to.
  • High out-degree users are highly active in reaching out to others.
  • These are not always the same users.

Extra: Detect communities in the network

user_graph_undirected <- as_undirected(user_graph, mode = "collapse")
comm <- cluster_louvain(user_graph_undirected)
length(comm)
[1] 6
modularity(comm)
[1] 0.2359954

Extra: Visualize communities

community_graph <- as_tbl_graph(user_graph_undirected) %>%
  activate(nodes) %>%
  mutate(community = as.factor(membership(comm)))

community_plot <- ggraph(community_graph, layout = "fr") +
  geom_edge_link(alpha = 0.15) +
  geom_node_point(aes(color = community), size = 3) +
  labs(title = "Communities in the Twitter Interaction Network") +
  theme_void()

print(community_plot)

Summary

In this tutorial, you learned how to:

  • construct a semantic network from bigrams
  • visualize relationships among words using igraph and ggraph
  • construct a social network from user mentions and replies
  • interpret common network measures such as degree, betweenness, density, and modularity

Network analysis is useful because it helps us move from isolated units of text or users to the broader relational structure that shapes communication.