Sentiment Analysis of Spotify, Amazon Music and Pandora user reviews

The datasets include almost 40,000 user reviews on Spotify, Amazon Music and Pandora collected from Trustpilot. Sentiment analysis is conducted to get an overview of users’ opinion on the three products and compare sentiment for each app in terms of subscription fee, UX/UI, music catalog and app integration with smart devices.

library(tidytext)
library(dplyr)
library(ggplot2)
library(stringr)
library(tidyr)
library(igraph)
library(ggraph)
library(tm)

Load the data

print("Spotify")

## [1] "Spotify"

str(spotify)

## 'data.frame':    7440 obs. of  3 variables:
##  $ ID    : int  1 2 3 4 5 6 7 8 9 10 ...
##  $ Date  : chr  "07-Oct-17" "04-Apr-17" "31-Jul-16" "07-Mar-17" ...
##  $ Review: chr  "Spotify is way, way, way, way, way, way better music app than Pandora. It doesn't limit you on the musicians yo"| __truncated__ "I love, love, love spotify !!!!! I'm 71 years and no very computer literate, if I can learn Spotify anyone can."| __truncated__ "I love this app. I put it on my phone first, but I had so much music it was taking up all the room on my phone."| __truncated__ "I love this app. I've tried them all (apple, amazon, pandora) and nothing compares. The suggestions are awesome"| __truncated__ ...

print("Amazon")

## [1] "Amazon"

str(amazon)

## 'data.frame':    12070 obs. of  3 variables:
##  $ ID    : int  1 2 3 4 5 6 7 8 9 10 ...
##  $ Date  : chr  "28-Mar-16" "30-Jul-17" "23-May-17" "04-May-17" ...
##  $ Review: chr  "love this music app. great selection. Better than Pandora. Use it on my Samsung Phone. Not happy, however, that"| __truncated__ "Terrific app." "Nice selection, I have it on a couple of devices, only problem is they're not synced together through amazon music." "It great piece to have tons of music to choose from! Loving it!" ...

print("Pandora")

## [1] "Pandora"

str(pandora)

## 'data.frame':    19780 obs. of  3 variables:
##  $ ï..ID : int  1 2 3 4 5 6 7 8 9 10 ...
##  $ Date  : chr  "11-Nov-17" "11-Dec-16" "12-Apr-16" "17-Nov-12" ...
##  $ Review: chr  "Love getting my music on all devices." "Love Pandora!! Can listen to my music on any device and in even in my car or on TV with DirectTV!" "i love the app and listen to Pandora often. Is there a voice command in my phone that would make changing stati"| __truncated__ "The pp just got an update that really helped smooth out a few problems or limitatikns in had, but its still got"| __truncated__ ...

colnames(pandora)[1] <- "ID"

Data prep

#Tokenize sentences
tidy_spotify <- spotify %>% unnest_tokens(word, Review)
tidy_amazon <- amazon %>% unnest_tokens(word, Review)
tidy_pandora <- pandora %>% unnest_tokens(word, Review)
stop_words <- stopwords("en")

cleaner <- function(x){
  filter(x, !word %in% stop_words)
}

tidy_spotify <- cleaner(tidy_spotify)
tidy_amazon <- cleaner(tidy_amazon)
tidy_pandora <- cleaner(tidy_pandora)

Sentiment analysis

Use 3 popular lexicons nrc, bing and afinn to assign numerical values to each word. The sum of scores given to all words (excluding stop words) in a review is used to classify whether a review is positive, neutral or negative.

#Get lexicons
get_sentiments(lexicon="nrc")

get_sentiments(lexicon="bing")

get_sentiments(lexicon="afinn")

#Join the lexicons
nrc_join <- function(x){
  x %>% inner_join(get_sentiments("nrc"), by = "word")
}

bing_join <- function(x){
  x %>% inner_join(get_sentiments("bing"), by = "word")
}

afinn_join <- function(x){
  x %>% inner_join(get_sentiments("afinn"), by = "word")
}

nrc_sentiment_spotify <- nrc_join(tidy_spotify)
nrc_sentiment_amazon <- nrc_join(tidy_amazon)
nrc_sentiment_pandora <- nrc_join(tidy_pandora)

bing_sentiment_spotify <- bing_join(tidy_spotify)
bing_sentiment_amazon <- bing_join(tidy_amazon)
bing_sentiment_pandora <- bing_join(tidy_pandora)

afinn_sentiment_spotify <- afinn_join(tidy_spotify)
afinn_sentiment_amazon <- afinn_join(tidy_amazon)
afinn_sentiment_pandora <- afinn_join(tidy_pandora)

Select only words with positive and negative sentiment from the nrc_sentiment dataframes and remove those with anger, anticipation, joy, etc. sentiment by subsetting:

#Aggregate review-level popularity information
#Subsetting function
subset_pn <- function(x){
  x[x$sentiment %in% c("positive", "negative"),]
}
nrc_sentiment_spotify <- subset_pn(nrc_sentiment_spotify)
nrc_sentiment_amazon <- subset_pn(nrc_sentiment_amazon)
nrc_sentiment_pandora <- subset_pn(nrc_sentiment_pandora)

#Aggregate review-level popularity information
nrc_aggregate <- function(x){
  x %>% select(ID, score) %>% 
    group_by(ID) %>% 
    summarise(nrc_score=sum(score))
}
bing_aggregate <- function(x){
  x %>% select(ID, score) %>% 
    group_by(ID) %>% 
    summarise(bing_score=sum(score))
}
afinn_aggregate <- function(x){
  x %>% select(ID, score) %>% 
    group_by(ID) %>% 
    summarise(afinn_score=sum(score))
}

Apply:

#Spotify
nrc_sentiment_spotify <- mutate(nrc_sentiment_spotify, score = ifelse(sentiment == "negative", -1, 1))
nrc_aggregate_spotify <- nrc_aggregate(nrc_sentiment_spotify)
bing_sentiment_spotify <- mutate(bing_sentiment_spotify, score = ifelse(sentiment == "negative", -1, 1))
bing_aggregate_spotify <- bing_aggregate(bing_sentiment_spotify)
afinn_aggregate_spotify <- afinn_aggregate(afinn_sentiment_spotify)
#Amazon
nrc_sentiment_amazon <- mutate(nrc_sentiment_amazon, score = ifelse(sentiment == "negative", -1, 1))
nrc_aggregate_amazon <- nrc_aggregate(nrc_sentiment_amazon)
bing_sentiment_amazon <- mutate(bing_sentiment_amazon, score = ifelse(sentiment == "negative", -1, 1))
bing_aggregate_amazon <- bing_aggregate(bing_sentiment_amazon)
afinn_aggregate_amazon <- afinn_aggregate(afinn_sentiment_amazon)
#Pandora
nrc_sentiment_pandora <- mutate(nrc_sentiment_pandora, score = ifelse(sentiment == "negative", -1, 1))
nrc_aggregate_pandora <- nrc_aggregate(nrc_sentiment_pandora)
bing_sentiment_pandora <- mutate(bing_sentiment_pandora, score = ifelse(sentiment == "negative", -1, 1))
bing_aggregate_pandora <- bing_aggregate(bing_sentiment_pandora)
afinn_aggregate_pandora <- afinn_aggregate(afinn_sentiment_pandora)

#Aggregate sentiment info to the original dataset
spotify_sent <- merge(x = spotify, y = nrc_aggregate_spotify, all.x = TRUE, by = "ID")
spotify_sent <- merge(x = spotify_sent, y = bing_aggregate_spotify, all.x = TRUE, by = "ID")
spotify_sent <- merge(x = spotify_sent, y = afinn_aggregate_spotify, all.x = TRUE, by = "ID")
spotify_sent[is.na(spotify_sent)] <- 0

amazon_sent <- merge(x = amazon, y = nrc_aggregate_amazon, all.x = TRUE, by = "ID")
amazon_sent <- merge(x = amazon_sent, y = bing_aggregate_amazon, all.x = TRUE, by = "ID")
amazon_sent <- merge(x = amazon_sent, y = afinn_aggregate_amazon, all.x = TRUE, by = "ID")
amazon_sent[is.na(amazon_sent)] <- 0

pandora_sent <- merge(x = pandora, y = nrc_aggregate_pandora, all.x = TRUE, by = "ID")
pandora_sent <- merge(x = pandora_sent, y = bing_aggregate_pandora, all.x = TRUE, by = "ID")
pandora_sent <- merge(x = pandora_sent, y = afinn_aggregate_pandora, all.x = TRUE, by = "ID")
pandora_sent[is.na(pandora_sent)] <- 0

Denote review sentiment:

#Denote review sentiment
##Spotify
  spotify_sent$afinn_judgement <- ifelse(spotify_sent$afinn_score < 0, "negative",
                                      ifelse(spotify_sent$afinn_score > 0, "positive", "neutral"))
  spotify_sent$bing_judgement <- ifelse(spotify_sent$bing_score < 0, "negative",
                                     ifelse(spotify_sent$bing_score > 0, "positive", "neutral"))
  spotify_sent$nrc_judgement <- ifelse(spotify_sent$nrc_score < 0, "negative",
                                    ifelse(spotify_sent$nrc_score > 0, "positive", "neutral"))
  rowSums(table(spotify_sent$bing_judgement, spotify_sent$bing_score))

## negative  neutral positive 
##      431     1178     5831

##Amazon
  amazon_sent$afinn_judgement <- ifelse(amazon_sent$afinn_score < 0, "negative",
                                         ifelse(amazon_sent$afinn_score > 0, "positive", "neutral"))
  amazon_sent$bing_judgement <- ifelse(amazon_sent$bing_score < 0, "negative",
                                        ifelse(amazon_sent$bing_score > 0, "positive", "neutral"))
  amazon_sent$nrc_judgement <- ifelse(amazon_sent$nrc_score < 0, "negative",
                                       ifelse(amazon_sent$nrc_score > 0, "positive", "neutral"))
  rowSums(table(amazon_sent$bing_judgement, amazon_sent$bing_score))

## negative  neutral positive 
##     1038     1889     9143

##Pandora
  pandora_sent$afinn_judgement <- ifelse(pandora_sent$afinn_score < 0, "negative",
                                         ifelse(pandora_sent$afinn_score > 0, "positive", "neutral"))
  pandora_sent$bing_judgement <- ifelse(pandora_sent$bing_score < 0, "negative",
                                        ifelse(pandora_sent$bing_score > 0, "positive", "neutral"))
  pandora_sent$nrc_judgement <- ifelse(pandora_sent$nrc_score < 0, "negative",
                                       ifelse(pandora_sent$nrc_score > 0, "positive", "neutral"))
  rowSums(table(pandora_sent$bing_judgement, pandora_sent$bing_score))

## negative  neutral positive 
##      964     2676    16140

Sentiment contribution

Which words contribute the most positive/negative sentiment towards the three apps?

Spotify:

bing_wc <- function(x){
  x %>%
    inner_join(get_sentiments("bing"), by = "word") %>%
    count(word, sentiment, sort = TRUE) %>%
    ungroup() %>%
    group_by(sentiment) %>%
    top_n(10) %>%
    ungroup() %>%
    mutate(word = reorder(word, n)) %>%
    ggplot(aes(word, n, fill = sentiment)) +
     geom_col(show.legend = FALSE) +
     facet_wrap(~sentiment, scales = "free_y") +
     labs(y = "Contribution to sentiment",
     x = NULL) +
     coord_flip()
}

See which words contribute the most to positive/negative sentiment on Spotify:

bing_wc(tidy_spotify)

## Selecting by n

Amazon:

bing_wc(tidy_amazon)

## Selecting by n

And Pandora:

bing_wc(tidy_pandora)

## Selecting by n

Word cloud

Word cloud of the most common words in each app’s review, with word size indicating frequency level:

library(wordcloud)

## Loading required package: RColorBrewer

c_wordcloud <- function(i){
  i %>%
  count(word) %>%
  with(wordcloud(word, n, max.words = 150, rot.per=0.1, scale = c(6,1), colors=brewer.pal(8, "Dark2")))
}

Spotify’s word cloud:

c_wordcloud(tidy_spotify)

Amazon’s word cloud:

c_wordcloud(tidy_amazon)

Pandora’s word cloud:

c_wordcloud(tidy_pandora)

Check review sentiment on specific topics

Price

How many percent of the reviews on price/subscription fee of each app are positive?

#Check price sentiment 
spotifyPrice <- grepl("cost| subscription| price| fee| premium| worth| money| pay| paid| pricing", spotify_sent$Review, ignore.case = TRUE)
sdf <- data.frame(sentiment = spotify_sent$bing_judgement, related_words = spotifyPrice)
sp = sum(sdf$related_words == "TRUE" & sdf$sentiment == "positive")/sum(sdf$related_words == "TRUE")*100
cat("Spotify price =", sp,"%")

## Spotify price = 75.9009 %

amazonPrice <- grepl("cost| subscription| price| fee| premium| worth| money| pay| paid| pricing", amazon_sent$Review, ignore.case = TRUE)
adf <- data.frame(sentiment = amazon_sent$bing_judgement, related_words = amazonPrice)
ap = sum(adf$related_words == "TRUE" & adf$sentiment == "positive")/sum(adf$related_words == "TRUE")*100
cat("Amazon price =", ap,"%")

## Amazon price = 69.80088 %

pandoraPrice <- grepl("cost| subscription| price| fee| premium| worth| money| pay| paid| pricing", pandora_sent$Review, ignore.case = TRUE)
pdf <- data.frame(sentiment = pandora_sent$bing_judgement, related_words = pandoraPrice)
pp = sum(pdf$related_words == "TRUE" & pdf$sentiment == "positive")/sum(pdf$related_words == "TRUE")*100
cat("Pandora price =", pp,"%")

## Pandora price = 79.85866 %

UX/UI

#Check UX/UI sentiment
spotifyUX <- grepl("functionality|UI|UX|experience|user experience|user interface|interface|usability|utility|app|glitch|device", spotify_sent$Review, ignore.case = TRUE)
sux <- data.frame(sentiment = spotify_sent$bing_judgement, related_words = spotifyUX)
su <- sum(sdf$related_words == "TRUE" & sdf$sentiment == "positive")/sum(sdf$related_words == "TRUE")*100
cat("Spotify UX =", su,"%")

## Spotify UX = 75.9009 %

amazonUX <- grepl("functionaliyu|UI|UX|experience|user experience|user interface|interface|usability|utility|app|glitch|device", amazon_sent$Review, ignore.case = TRUE)
aux <- data.frame(sentiment = amazon_sent$bing_judgement, related_words = amazonUX)
au <- sum(aux$related_words == "TRUE" & aux$sentiment == "positive")/sum(aux$related_words == "TRUE")*100
cat("Amazon UX =", au,"%")

## Amazon UX = 70.12021 %

pandoraUX <- grepl("functionality|UI|UX|experience|user experience|user interface|interface|usability|utility|app|glitch|device", pandora_sent$Review, ignore.case = TRUE)
pux <- data.frame(sentiment = pandora_sent$bing_judgement, related_words = pandoraUX)
pu <- sum(pux$related_words == "TRUE" & pux$sentiment == "positive")/sum(pux$related_words == "TRUE")*100
cat("Pandora UX", pu,"%")

## Pandora UX 82.72777 %

Music catalog

#Check music catalog 
##Spotify
spotifyMusic <- grepl("music selection| selection| music| customized| playlist| recommendation| AI| discovery| variety| song| songs| library| broad| wide| range| catalogue| list| singers", spotify_sent$Review, ignore.case = TRUE)
sm <- data.frame(sentiment = spotify_sent$bing_judgement, related_words = spotifyMusic)
smc <- sum(sm$related_words == "TRUE" & sm$sentiment == "positive")/sum(sm$related_words == "TRUE")*100
cat("Spotify catalog =", smc,"%")

## Spotify catalog = 83.78454 %

##Amzon
amazonMusic <- grepl("music selection| selection| music| customized| playlist| recommendation| AI| discovery| variety| song| songs| library| broad| wide| range| catalogue| list| singers", amazon_sent$Review, ignore.case = TRUE)
amusic <- data.frame(sentiment = amazon_sent$bing_judgement, related_words = amazonMusic)
amc <- sum(amusic$related_words == "TRUE" & amusic$sentiment == "positive")/sum(amusic$related_words == "TRUE")*100
cat("Amazon catalog =", amc,"%")

## Amazon catalog = 76.60958 %

##Pandora
pandoraMusic <- pandoraMusic <- grepl("music selection| selection| music| customized| playlist| recommendation| AI| discovery| variety| song| songs| library| broad| wide| range| catalogue| list| singers", pandora_sent$Review, ignore.case = TRUE)
pmusic <- data.frame(sentiment = pandora_sent$bing_judgement, related_words = pandoraMusic)
pmc <- sum(pmusic$related_words == "TRUE" & pmusic$sentiment == "positive")/sum(pmusic$related_words == "TRUE")*100
cat("Pandora catalog =", pmc,"%")

## Pandora catalog = 86.57468 %

App’s integration with smart devices

#Check integration 
spotifyIn <- grepl("device| connection| connect| speaker| smart speaker| alexa| google home| siri| echo| integrate| integration| seamless| across| sync| synchronize| synchronization| bluetooth| strong| car| headphone| earphone| phone| airpod", spotify_sent$Review, ignore.case = TRUE)
si <- data.frame(sentiment = spotify_sent$bing_judgement, related_words = spotifyIn)
s <- sum(si$related_words == "TRUE" & si$sentiment == "positive")/sum(si$related_words == "TRUE")*100
cat("Spotify integration =", s,"%")

## Spotify integration = 74.68619 %

amazonIn <- grepl("device| connection| connect| speaker| smart speaker| alexa| google home| siri| echo| integrate| integration| seamless| across| sync| synchronize| synchronization| bluetooth| strong| car| headphone| earphone| phone| airpod", amazon_sent$Review, ignore.case = TRUE)
ai <- data.frame(sentiment = amazon_sent$bing_judgement, related_words = amazonIn)
a <- sum(ai$related_words == "TRUE" & ai$sentiment == "positive")/sum(ai$related_words == "TRUE")*100
cat("Amazon integration =", a,"%")

## Amazon integration = 66.28849 %

pandoraIn <- grepl("device| connection| connect| speaker| smart speaker| alexa| google home| siri| echo| integrate| integration| seamless| across| sync| synchronize| synchronization| bluetooth| strong| car| headphone| earphone| phone| airpod", pandora_sent$Review, ignore.case = TRUE)
pi <- data.frame(sentiment = pandora_sent$bing_judgement, related_words = pandoraIn)
p <- sum(pi$related_words == "TRUE" & pi$sentiment == "positive")/sum(pi$related_words == "TRUE")*100
cat("Pandora integration =", p, "%")

## Pandora integration = 81.35304 %

N-gram visualization

N-gram is a useful technique to see which words often go together in the document, so that we can make more sense of the text’s content. Here we will start with bigrams (pairs of words).

#Word clustering using bigrams
bi_graph <- subset(spotify_sent, bing_judgement == "positive")
app_bigram <- bi_graph %>% unnest_tokens(bigram, Review, token = "ngrams", n = 2)

##Separate 
bigrams_separated <- app_bigram %>% 
    separate(bigram, c("word1", "word2"), sep = " ")

##Filter 
bigrams_filtered <- bigrams_separated %>%
    filter(!word1 %in% stop_words) %>%
    filter(!word2 %in% stop_words)

##New bigram counts:
bigram_counts <- bigrams_filtered %>%
    count(word1, word2, sort = TRUE)
  
##Unite
bigrams_united <- bigrams_filtered %>%
    unite(bigram, word1, word2, sep = " ")

##Filter for only relatively common combinations
bigram_graph <- bigram_counts %>%
    filter(n > 10) %>%
    graph_from_data_frame() 

## Warning in graph_from_data_frame(.): In `d' `NA' elements were replaced
## with string "NA"

bigram_graph

## IGRAPH 0928570 DN-- 157 190 -- 
## + attr: name (v/c), n (e/n)
## + edges from 0928570 (vertex names):
##  [1] music   ->app       NA      ->NA        great   ->app      
##  [4] best    ->music     can     ->listen    love    ->spotify  
##  [7] great   ->music     good    ->app       free    ->music    
## [10] app     ->ever      favorite->music     kindle  ->fire     
## [13] favorite->songs     love    ->love      works   ->great    
## [16] free    ->version   best    ->app       new     ->music    
## [19] awesome ->app       can     ->find      good    ->music    
## [22] highly  ->recommend much    ->better    every   ->song     
## + ... omitted several edges

##void-themed graph 
  a <- grid::arrow(type = "closed", length = unit(0.15, "inches"))
  ggraph(bigram_graph, layout = "fr") +
  geom_edge_link(show.legend = FALSE,
                   end_cap = circle(.07, 'inches')) +
  geom_node_point(color = "seagreen3", size = 3) +
  geom_node_text(aes(label = name), vjust = 1, hjust = 1) +
  theme_void() + 
  labs(title="Spotify bigrams")

With Amazon:

#Word clustering using bigrams
library(tidyr)
bi_graph <- subset(amazon_sent, bing_judgement == "positive")
app_bigram <- bi_graph %>% unnest_tokens(bigram, Review, token = "ngrams", n = 2) 
app_bigram %>%
  count(bigram, sort = TRUE)

##Separate 
bigrams_separated <- app_bigram %>% 
    separate(bigram, c("word1", "word2"), sep = " ")

##Filter 
bigrams_filtered <- bigrams_separated %>%
    filter(!word1 %in% stop_words) %>%
    filter(!word2 %in% stop_words)
##New bigram counts:
bigram_counts <- bigrams_filtered %>% 
    count(word1, word2, sort = TRUE)
  
##Unite
bigrams_united <- bigrams_filtered %>%
    unite(bigram, word1, word2, sep = " ")

##Filter the most common bigrams
bigram_graph <- bigram_counts %>%
    filter(n > 20) %>%
    graph_from_data_frame()

##void-themed graph 
a <- grid::arrow(type = "closed", length = unit(0.15, "inches"))
ggraph(bigram_graph, layout = "fr") +
  geom_edge_link(show.legend = FALSE,
                   end_cap = circle(.07, 'inches')) +
  geom_node_point(color = "orange", size = 3) +
  geom_node_text(aes(label = name), vjust = 1, hjust = 1) +
  theme_void() + 
  labs(title="Amazon bigrams")

And Pandora:

#Word clustering using bigrams
library(tidyr)
bi_graph <- subset(pandora_sent, bing_judgement == "positive")
app_bigram <- bi_graph %>% unnest_tokens(bigram, Review, token = "ngrams", n = 2) 
app_bigram %>%
  count(bigram, sort = TRUE)

##Separate 
bigrams_separated <- app_bigram %>% 
    separate(bigram, c("word1", "word2"), sep = " ")

##Filter 
bigrams_filtered <- bigrams_separated %>%
    filter(!word1 %in% stop_words) %>%
    filter(!word2 %in% stop_words)
##New bigram counts:
bigram_counts <- bigrams_filtered %>% 
    count(word1, word2, sort = TRUE)
  
##Unite
bigrams_united <- bigrams_filtered %>%
    unite(bigram, word1, word2, sep = " ")

##Filter the most common bigrams
bigram_graph <- bigram_counts %>%
    filter(n > 25) %>%
    graph_from_data_frame()

##void-themed graph 
a <- grid::arrow(type = "closed", length = unit(0.15, "inches"))
ggraph(bigram_graph, layout = "fr") +
  geom_edge_link(show.legend = FALSE,
                   end_cap = circle(.07, 'inches')) +
  geom_node_point(color = "skyblue2", size = 3) +
  geom_node_text(aes(label = name), vjust = 1, hjust = 1) +
  theme_void() + 
  labs(title="Pandora bigrams")

The bigram plots above could help us see the links between the most common words, better understand review contexts and explore notable topics for further analysis.