作者|Audhi Aprilliant
编译|VK
来源|Towards Datas Science

概述

对于这个项目，我们在2019年5月28-29日通过爬虫来使用Twitter的原始数据。此外，数据是CSV格式（逗号分隔），可以在这里下载。

https://github.com/audhiaprilliant/Indonesia-Public-Election-Twitter-Sentiment-Analysis/tree/master/Datasets

它涉及两个主题，一个是包含关键字“Joko Widodo”的Joko Widodo的数据，另一个是带有关键字“Prabowo Subianto”的Prabowo Subianto的数据。其中包括几个变量和信息，以确定用户情绪。实际上，数据有16个变量或属性和1000多个观察值。表1列出了一些变量。

# 导入库
library(ggplot2)
library(lubridate)

# 加载Joko Widodo的数据
data.jokowi.df = read.csv(file = \'data-joko-widodo.csv\',
                          header = TRUE,
                          sep = \',\')
senti.jokowi = read.csv(file = \'sentiment-joko-widodo.csv\',
                        header = TRUE,
                        sep = \',\')
                        
# 加载Prabowo Subianto的数据
data.prabowo.df = read.csv(file = \'data-prabowo-subianto.csv\',
                           header = TRUE,
                           sep = \',\')
senti.prabowo = read.csv(file = \'sentiment-prabowo-subianto.csv\',
                         header = TRUE,
                         sep = \',\')

数据可视化

数据探索旨在从Twitter数据中获取任何信息。应该指出的是，数据已经进行了文本预处理。我们对那些被认为是很有趣的变量进行探索。。

# TWEETS的条形图-JOKO WIDODO
data.jokowi.df$created = ymd_hms(data.jokowi.df$created,
                                 tz = \'Asia/Jakarta\')
# 另一种制作“date”和“hour”变量的方法
data.jokowi.df$date = date(data.jokowi.df$created)
data.jokowi.df$hour = hour(data.jokowi.df$created)
# 日期2019-05-29
data.jokowi.date1 = subset(x = data.jokowi.df,
                           date == \'2019-05-29\')
data.hour.date1 = data.frame(table(data.jokowi.date1$hour))
colnames(data.hour.date1) = c(\'Hour\',\'Total.Tweets\')
# 创建数据可视化
ggplot(data.hour.date1)+
  geom_bar(aes(x = Hour,
               y = Total.Tweets,
               fill = I(\'blue\')),
           stat = \'identity\',
           alpha = 0.75,
           show.legend = FALSE)+
  geom_hline(yintercept = mean(data.hour.date1$Total.Tweets),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average:\',
ceiling(mean(data.hour.date1$Total.Tweets)),
                              \'Tweets per hour\'),
                x = 8,
                y = mean(data.hour.date1$Total.Tweets)+20),
            hjust = \'left\',
            size = 4)+
  labs(title = \'Total Tweets per Hours - Joko Widodo\',
       subtitle = \'28 May 2019\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Time of Day\')+
  ylab(\'Total Tweets\')+
  scale_fill_brewer(palette = \'Dark2\')+
  theme_bw()
  
# TWEETS的条形图-PRABOWO SUBIANTO
data.prabowo.df$created = ymd_hms(data.prabowo.df$created,
                                  tz = \'Asia/Jakarta\')
                                  
# 另一种制作“date”和“hour”变量的方法
data.prabowo.df$date = date(data.prabowo.df$created)
data.prabowo.df$hour = hour(data.prabowo.df$created)

# 日期2019-05-28
data.prabowo.date1 = subset(x = data.prabowo.df,
                            date == \'2019-05-28\')
data.hour.date1 = data.frame(table(data.prabowo.date1$hour))
colnames(data.hour.date1) = c(\'Hour\',\'Total.Tweets\')

# 日期 2019-05-29
data.prabowo.date2 = subset(x = data.prabowo.df,
                            date == \'2019-05-29\')
data.hour.date2 = data.frame(table(data.prabowo.date2$hour))
colnames(data.hour.date2) = c(\'Hour\',\'Total.Tweets\')
data.hour.date3 = rbind(data.hour.date1,data.hour.date2)
data.hour.date3$Date = c(rep(x = \'2019-05-28\',
                             len = nrow(data.hour.date1)),
                         rep(x = \'2019-05-29\',
                             len = nrow(data.hour.date2)))
data.hour.date3$Labels = c(letters,\'A\',\'B\')
data.hour.date3$Hour = as.character(data.hour.date3$Hour)
data.hour.date3$Hour = as.numeric(data.hour.date3$Hour)

# 数据预处理
for (i in 1:nrow(data.hour.date3)) {
  if (i%%2 == 0) {
    data.hour.date3[i,\'Hour\'] = \'\'
  }
  if (i%%2 == 1) {
    data.hour.date3[i,\'Hour\'] = data.hour.date3[i,\'Hour\']
  }
}
data.hour.date3$Hour = as.factor(data.hour.date3$Hour)

# 数据可视化
ggplot(data.hour.date3)+
  geom_bar(aes(x = Labels,
               y = Total.Tweets,
               fill = Date),
           stat = \'identity\',
           alpha = 0.75,
           show.legend = TRUE)+
  geom_hline(yintercept = mean(data.hour.date3$Total.Tweets),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average:\',
ceiling(mean(data.hour.date3$Total.Tweets)),
                              \'Tweets per hour\'),
                x = 5,
                y = mean(data.hour.date3$Total.Tweets)+6),
            hjust = \'left\',
            size = 3.8)+
  scale_x_discrete(limits = data.hour.date3$Labels,
                   labels = data.hour.date3$Hour)+
  labs(title = \'Total Tweets per Hours - Prabowo Subianto\',
       subtitle = \'28 - 29 May 2019\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Time of Day\')+
  ylab(\'Total Tweets\')+
  ylim(c(0,100))+
  theme_bw()+
  theme(legend.position = \'bottom\',
        legend.title = element_blank())+
  scale_fill_brewer(palette = \'Dark2\')

根据图1，我们可以得出结论，通过数据抓取（关键字“Jokow Widodo”和“Prabowo Subianto”）得到的tweet数量并不相似，即使在同一日期。

例如，在图1（左）中，从视觉上看，对于关键字为“Joko Widodo”的推文，仅在2019年5月28日03:00–17:00 WIB期间获得。而在图1（右图）中，我们得出的结论是，在2019年5月28日至29日12:00-23:59 WIB（2019年5月28日）和00:00-15:00 WIB（2019年5月29日）期间获得的关键词为“Prabowo Subianto”的推文。

# 2019-05-28的推特
ggplot(data.hour.date1)+
  geom_bar(aes(x = Hour,
               y = Total.Tweets,
               fill = I(\'red\')),
           stat = \'identity\',
           alpha = 0.75,
           show.legend = FALSE)+
  geom_hline(yintercept = mean(data.hour.date1$Total.Tweets),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average:\',
ceiling(mean(data.hour.date1$Total.Tweets)),
                              \'Tweets per hour\'),
                x = 6.5,
                y = mean(data.hour.date1$Total.Tweets)+5),
            hjust = \'left\',
            size = 4)+
  labs(title = \'Total Tweets per Hours - Prabowo Subianto\',
       subtitle = \'28 May 2019\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Time of Day\')+
  ylab(\'Total Tweets\')+
  ylim(c(0,100))+
  theme_bw()+
  scale_fill_brewer(palette = \'Dark2\')
  
# 2019-05-29的推特
ggplot(data.hour.date2)+
  geom_bar(aes(x = Hour,
               y = Total.Tweets,
               fill = I(\'red\')),
           stat = \'identity\',
           alpha = 0.75,
           show.legend = FALSE)+
  geom_hline(yintercept = mean(data.hour.date2$Total.Tweets),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average:\',
ceiling(mean(data.hour.date2$Total.Tweets)),
                              \'Tweets per hour\'),
                x = 1,
                y = mean(data.hour.date2$Total.Tweets)+6),
            hjust = \'left\',
            size = 4)+
  labs(title = \'Total Tweets per Hours - Prabowo Subianto\',
       subtitle = \'29 May 2019\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Time of Day\')+
  ylab(\'Total Tweets\')+
  ylim(c(0,100))+
  theme_bw()+
  scale_fill_brewer(palette = \'Dark2\')

根据图2，我们得到了使用关键字“Joko Widodo”和“Prabowo Subianto”的用户之间的显著差异。关键词为“Joko Widodo”的tweet在某个特定时间（07:00–09:00 WIB）谈论Joko Widodo往往非常激烈，08:00 WIB的tweet数量最多。它有348条推文。然而，在2019年5月28日至29日期间，关键词为“Prabowo Subianto”的推文往往会不断地谈论Prabowo Subianto。2019年5月28日至29日，每小时上传关键词为“Prabowo Subianto”的推文平均为36条。

# JOKO WIDODO
df.score.1 = subset(senti.jokowi,class == c(\'Negative\',\'Positive\'))
colnames(df.score.1) = c(\'Score\',\'Text\',\'Sentiment\')
# Data viz
ggplot(df.score.1)+
  geom_density(aes(x = Score,
                   fill = Sentiment),
               alpha = 0.75)+
  xlim(c(-11,11))+
  labs(title = \'Density Plot of Sentiment Scores\',
       subtitle = \'Joko Widodo\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Score\')+ 
  ylab(\'Density\')+
  theme_bw()+
  scale_fill_brewer(palette = \'Dark2\')+
  theme(legend.position = \'bottom\',
        legend.title = element_blank())
        
# PRABOWO SUBIANTO
df.score.2 = subset(senti.prabowo,class == c(\'Negative\',\'Positive\'))
colnames(df.score.2) = c(\'Score\',\'Text\',\'Sentiment\')
ggplot(df.score.2)+
  geom_density(aes(x = Score,
                   fill = Sentiment),
               alpha = 0.75)+
  xlim(c(-11,11))+
  labs(title = \'Density Plot of Sentiment Scores\',
       subtitle = \'Prabowo Subianto\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Density\')+ 
  ylab(\'Score\')+
  theme_bw()+
  scale_fill_brewer(palette = \'Dark2\')+
  theme(legend.position = \'bottom\',
        legend.title = element_blank())

图3是2019年5月28日至29日以“Joko Widodo”和“Prabowo Subianto”为关键词的多条推文的条形图。由图3(左)可以得出，Twitter用户在19:00-23:59 WIB上谈论Prabowo Subianto的频率较低。这是由于印尼人的休息时间造成的。然而，这些带有主题的推文总是在午夜更新，因为有的用户居住在国外，有的用户仍然活跃。然后，用户在04:00 WIB开始活动，在07:00 WIB达到高峰，然后下降，直到12:00 WIB再次上升。

# JOKO WIDODO
df.senti.score.1 = data.frame(table(senti.jokowi$score))
colnames(df.senti.score.1) = c(\'Score\',\'Freq\')
# 数据预处理
df.senti.score.1$Score = as.character(df.senti.score.1$Score)
df.senti.score.1$Score = as.numeric(df.senti.score.1$Score)
Score1 = df.senti.score.1$Score
sign(df.senti.score.1[1,1])
for (i in 1:nrow(df.senti.score.1)) {
  sign.row = sign(df.senti.score.1[i,\'Score\'])
  for (j in 1:ncol(df.senti.score.1)) {
    df.senti.score.1[i,j] = df.senti.score.1[i,j] * sign.row
  }
}
df.senti.score.1$Label = c(letters[1:nrow(df.senti.score.1)])
df.senti.score.1$Sentiment = ifelse(df.senti.score.1$Freq < 0,
                                    \'Negative\',\'Positive\')
df.senti.score.1$Score1 = Score1
# 数据可视化
ggplot(df.senti.score.1)+
  geom_bar(aes(x = Label,
               y = Freq,
               fill = Sentiment),
           stat = \'identity\',
           show.legend = FALSE)+
  # 积极情感
  geom_hline(yintercept = mean(abs(df.senti.score.1[which(df.senti.score.1$Sentiment == \'Positive\'),\'Freq\'])),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average Freq:\',
ceiling(mean(abs(df.senti.score.1[which(df.senti.score.1$Sentiment == \'Positive\'),\'Freq\'])))),
                x = 10,
                y = mean(abs(df.senti.score.1[which(df.senti.score.1$Sentiment == \'Positive\'),\'Freq\']))+30),
            hjust = \'right\',
            size = 4)+
  # 消极情感
  geom_hline(yintercept = mean(df.senti.score.1[which(df.senti.score.1$Sentiment == \'Negative\'),\'Freq\']),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average Freq:\',
ceiling(mean(abs(df.senti.score.1[which(df.senti.score.1$Sentiment == \'Negative\'),\'Freq\'])))),
                x = 5,
                y = mean(df.senti.score.1[which(df.senti.score.1$Sentiment == \'Negative\'),\'Freq\'])-15),
            hjust = \'left\',
            size = 4)+
  labs(title = \'Barplot of Sentiments\',
       subtitle = \'Joko Widodo\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Score\')+
  scale_x_discrete(limits = df.senti.score.1$Label,
                   labels = df.senti.score.1$Score1)+
  theme_bw()+
  scale_fill_brewer(palette = \'Dark2\')

# PRABOWO SUBIANTO
df.senti.score.2 = data.frame(table(senti.prabowo$score))
colnames(df.senti.score.2) = c(\'Score\',\'Freq\')
# 数据预处理
df.senti.score.2$Score = as.character(df.senti.score.2$Score)
df.senti.score.2$Score = as.numeric(df.senti.score.2$Score)
Score2 = df.senti.score.2$Score
sign(df.senti.score.2[1,1])
for (i in 1:nrow(df.senti.score.2)) {
  sign.row = sign(df.senti.score.2[i,\'Score\'])
  for (j in 1:ncol(df.senti.score.2)) {
    df.senti.score.2[i,j] = df.senti.score.2[i,j] * sign.row
  }
}
df.senti.score.2$Label = c(letters[1:nrow(df.senti.score.2)])
df.senti.score.2$Sentiment = ifelse(df.senti.score.2$Freq < 0,
                                    \'Negative\',\'Positive\')
df.senti.score.2$Score1 = Score2
# 数据可视化
ggplot(df.senti.score.2)+
  geom_bar(aes(x = Label,
               y = Freq,
               fill = Sentiment),
           stat = \'identity\',
           show.legend = FALSE)+
  # 积极情感
  geom_hline(yintercept = mean(abs(df.senti.score.2[which(df.senti.score.2$Sentiment == \'Positive\'),\'Freq\'])),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average Freq:\',
ceiling(mean(abs(df.senti.score.2[which(df.senti.score.2$Sentiment == \'Positive\'),\'Freq\'])))),
                x = 11,
                y = mean(abs(df.senti.score.2[which(df.senti.score.2$Sentiment == \'Positive\'),\'Freq\']))+20),
            hjust = \'right\',
            size = 4)+
  # 消极情感
  geom_hline(yintercept = mean(df.senti.score.2[which(df.senti.score.2$Sentiment == \'Negative\'),\'Freq\']),
             col = I(\'black\'),
             size = 1)+
  geom_text(aes(fontface = \'italic\',
                label = paste(\'Average Freq:\',
ceiling(mean(abs(df.senti.score.2[which(df.senti.score.2$Sentiment == \'Negative\'),\'Freq\'])))),
                x = 9,
                y = mean(df.senti.score.2[which(df.senti.score.2$Sentiment == \'Negative\'),\'Freq\'])-10),
            hjust = \'left\',
            size = 4)+
  labs(title = \'Barplot of Sentiments\',
       subtitle = \'Prabowo Subianto\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlab(\'Score\')+
  scale_x_discrete(limits = df.senti.score.2$Label,
                   labels = df.senti.score.2$Score1)+
  theme_bw()+
  scale_fill_brewer(palette = \'Dark2\')

图4是包含关键字“Joko Widodo”和“Prabowo Subianto”的情感得分密度图。tweets的得分是由组成tweets的词根的平均得分得到的。因此，它的分数是针对每个词根给出的，其值介于-10到10之间。如果分数越小，那么微博中的负面情绪就越多，反之亦然。根据图4（左），可以得出结论，包含关键字“Joko Widodo”的推文的负面情绪在-10到-1之间，中间得分为-4。它也适用于积极的情绪（当然，有一个积极的分数）。根据图4（左）中的密度图，我们发现积极情绪的得分具有相当小的方差。因此，我们得出结论，对包含关键词“Joko Widodo”的微博的积极情绪并不是太多样化。

图4（右）显示了包含关键字“Prabowo Subianto”的情感得分密度图。它与图4（左）不同，因为图4（右）上的负面情绪在-8到-1之间。这意味着tweets没有太多负面情绪（tweets有负面情绪，但不够高）。此外，负面情绪得分的分布在4和1之间有两个峰值。然而，积极情绪从1到10不等。与图4（左）相比，图4（右）的积极情绪具有较高的方差，在3和10范围内有两个峰值。这表明，包含关键词“Prabowo Subianto”的微博具有很高的积极情绪。

# JOKO WIDODO
df.senti.3 = as.data.frame(table(senti.jokowi$class))
colnames(df.senti.3) = c(\'Sentiment\',\'Freq\')
# 数据预处理
df.pie.1 = df.senti.3
df.pie.1$Prop = df.pie.1$Freq/sum(df.pie.1$Freq)
df.pie.1 = df.pie.1 %>%
  arrange(desc(Sentiment)) %>%
  mutate(lab.ypos = cumsum(Prop) - 0.5*Prop)
# 数据可视化
ggplot(df.pie.1,
       aes(x = 2,
           y = Prop,
           fill = Sentiment))+
  geom_bar(stat = \'identity\',
           col = \'white\',
           alpha = 0.75,
           show.legend = TRUE)+
  coord_polar(theta = \'y\', 
              start = 0)+
  geom_text(aes(y = lab.ypos,
                label = Prop),
            color = \'white\',
            fontface = \'italic\',
            size = 4)+
  labs(title = \'Piechart of Sentiments\',
       subtitle = \'Joko Widodo\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlim(c(0.5,2.5))+
  theme_void()+
  scale_fill_brewer(palette = \'Dark2\')+
  theme(legend.title = element_blank(),
        legend.position = \'right\')
        
# PRABOWO SUBIANTO
df.senti.4 = as.data.frame(table(senti.prabowo$class))
colnames(df.senti.4) = c(\'Sentiment\',\'Freq\')
# 数据预处理
df.pie.2 = df.senti.4
df.pie.2$Prop = df.pie.2$Freq/sum(df.pie.2$Freq)
df.pie.2 = df.pie.2 %>%
  arrange(desc(Sentiment)) %>%
  mutate(lab.ypos = cumsum(Prop) - 0.5*Prop)
# 数据可视化
ggplot(df.pie.2,
       aes(x = 2,
           y = Prop,
           fill = Sentiment))+
  geom_bar(stat = \'identity\',
           col = \'white\',
           alpha = 0.75,
           show.legend = TRUE)+
  coord_polar(theta = \'y\', 
              start = 0)+
  geom_text(aes(y = lab.ypos,
                label = Prop),
            color = \'white\',
            fontface = \'italic\',
            size = 4)+
  labs(title = \'Piechart of Sentiments\',
       subtitle = \'Prabowo Subianto\',
       caption = \'Twitter Crawling 28 - 29 May 2019\')+
  xlim(c(0.5,2.5))+
  theme_void()+
  scale_fill_brewer(palette = \'Dark2\')+
  theme(legend.title = element_blank(),
        legend.position = \'right\')

图5是推特的情绪得分汇总，这些微博被分为负面情绪、中性情绪和积极情绪。消极情绪是指得分低于零的情绪，中性是指分数等于零的情绪，积极情绪得分大于零。从图5可以看出，关键字为“Joko Widodo”的微博的负面情绪百分比低于关键字为“Prabowo Subianto”的tweet。有6.3%的差异。研究还发现，与关键词为Prabowo Subianto的微博相比，包含关键词“Joko Widodo”的微博具有更高的中性情绪和积极情绪。通过piechart的研究发现，与关键字为“Prabowo Subianto”的tweet相比，带有关键字“Joko Widodo”的tweet倾向于拥有更高比例的积极情绪。但是通过密度图发现，积极和消极情绪得分的分布表明，与“Joko Widodo”相比，包含关键字“Prabowo Subianto”的微博往往具有更高的情绪得分。它必须进行进一步的分析。

图6显示了用户在2019年5月28-29日经常上传的tweet（关键词“Joko Widodo”和“Prabowo Subianto”）中的术语或单词。通过这个WordCloud可视化，可以找到热门话题，这些话题都是针对关键词进行讨论的。对于包含关键词“Joko Widodo”的tweet，我们发现术语“tuang”、“petisi”、“negara”、“aman”和“nusantara”是前五名，每个tweet出现的次数最多。然而，包含关键词“Joko Widodo”的tweet发现，“Prabowo”、“Subianto”、“kriminalisasi”、“selamat”和“dubai”是每个tweet中出现次数最多的前五个词。这间接地显示了以关键字“Prabowo Subianto”上传的tweet的模式，即：几乎可以肯定的是，每个上传的tweet都直接包含“Prabowo Subianto”的名称，而不是通过提及（@）。这是因为，在文本预处理中，提到（@）已被删除。

可以前往我的GitHub repo查找代码：https://github.com/audhiaprilliant/Indonesia-Public-Election-Twitter-Sentiment-Analysis

参考引用

[1] K. Borau, C. Ullrich, J. Feng, R. Shen. Microblogging for Language Learning: Using Twitter to Train Communicative and Cultural Competence (2009), Advances in Web-Based Learning — ICWL 2009, 8th International Conference, Aachen, Germany, August 19–21, 2009.

原文链接：https://towardsdatascience.com/twitter-data-visualization-fb4f45b63728

欢迎关注磐创AI博客站：
http://panchuang.net/

sklearn机器学习中文官方文档：
http://sklearn123.com/

欢迎关注磐创博客资源汇总站：
http://docs.panchuang.net/