Dancing Through the Years: A Data-Driven Look at Taylor Swift’s Music

Looking at the W. Jake Thompson’s curated data-set of Taylor Swift songs as a part of #TidyTuesday (Oct 10, 2023)

Step 1: Data Import

Getting the data from TidyTuesday Retrieve the data originally from the taylor R package from W. Jake Thompson is a curated data set of Taylor Swift songs, including lyrics and audio characteristics. The data comes from Genius and the Spotify API.

library(tidyverse)       # Data Wrangling and Visualization
library(visdat)          # View data in Exploratory Data Analysis
library(gganimate)       # For animation
library(transformr)      # to smoothly animate polygons and paths

# Using Option 2: Read data directly from GitHub

taylor_album_songs <- readr::read_csv('https://raw.githubusercontent.com/rfordatascience/tidytuesday/master/data/2023/2023-10-17/taylor_album_songs.csv')
taylor_all_songs <- readr::read_csv('https://raw.githubusercontent.com/rfordatascience/tidytuesday/master/data/2023/2023-10-17/taylor_all_songs.csv')
taylor_albums <- readr::read_csv('https://raw.githubusercontent.com/rfordatascience/tidytuesday/master/data/2023/2023-10-17/taylor_albums.csv')

Step 2: Some Exploratory Data Analysis

# Since all songs of Taylor Swift occur in taylor_all_songs, let us
# focus on that data set only for now
taylor_album_songs |> 
  anti_join(taylor_all_songs)

# Seeing the number of distinct values for each variable
taylor_all_songs |> 
  summarise(across(.cols = everything(),
                .fns = n_distinct)) |> 
  pivot_longer(cols = everything(),
               names_to = "Variable",
               values_to = "n_distinct")

Using a popular function vis_dat() to see the structure of the data: –

# Vis_dat the data
taylor_all_songs |> 
  vis_dat()

And, seeing the change in different song characteristics over time: –

# We see the patterns over time for different variables of her 
# songs to see any distinct patterns
taylor_all_songs |>
  select(album_name, track_name, track_release,
         danceability:duration_ms) |> 
  pivot_longer(cols = -c(album_name, track_name, track_release),
               names_to = "indicator",
               values_to = "value") |> 
  ggplot(aes(x = track_release,
             y = value)) +
  geom_point() +
  geom_smooth() +
  facet_wrap(~ indicator, scales = "free") +
  theme_classic()

Creating a static graph which we will animate later, and setting the span parameter for loess smoother: –

#define span to use
span_taylor = 0.75

# Take the taylor_all_songs data frame and select specific columns:
taylor_all_songs |> 
  select(album_name,
         track_release,
         danceability, 
         acousticness) |> 

# Pivot the selected columns into a longer format with "indicators" and "values" columns:
  pivot_longer(cols = -c(album_name, track_release),
               names_to = "indicators",
               values_to = "values") |>

# Create a ggplot visualization, setting aesthetics and geometries:
  ggplot(aes(x = track_release,
             y = values,
             col = indicators,
             label = indicators)) +

# Add jittered points to the plot with specified width, height, and alpha:
  geom_jitter(width = 20, 
              height = 0.001, 
              alpha = 0.2) +

# Add a smoothed line to the plot with specified span, se, and alpha:
  geom_smooth(span = span_taylor, 
              se = FALSE,
              alpha = 0.6,
              lwd = 1.2) +

# Add text labels to the plot, referencing data from taylor_albums:
  geom_text(data = taylor_albums, 
             mapping = aes(x = album_release,
                           y = 0,
                           label = album_name),
             col = "black",
             angle = 90, 
            hjust = "left") +

# Apply a minimal theme to the plot:
  theme_minimal() +

# Customize the x-axis labels using breaks and formatted labels:
  scale_x_continuous(breaks = taylor_albums$album_release,
                     labels = format(taylor_albums$album_release, 
                                     "%b %Y")) +

# Using color palettes from the tayloRswift package for Taylor Swift's albums:
  tayloRswift::scale_color_taylor(palette = "lover") +

# Add labels and customize the appearance of the plot:
  labs(x = NULL,
       y = "Spotify App Score for songs",
       color = NULL) +

# Further customize the appearance of the plot using theme settings:
  theme(axis.text.x = element_text(angle = 90),
        panel.grid.minor.x = element_blank(),
        panel.grid.minor.y = element_blank(), 
        legend.position = "bottom")

Step 3: Creating an animated plot

# creating a loess predictor variable
df <- taylor_all_songs |> 
  select(album_name,
         track_release,
         danceability, 
         acousticness) |> 
  drop_na()
  
gganim <- df |> 
  mutate(
    smooth_dance = predict(loess(danceability ~ as.numeric(track_release), 
                                data = df,
                                span = span_taylor)),
    smooth_acous = predict(loess(acousticness ~ as.numeric(track_release), 
                                data = df,
                                span = span_taylor))
  ) |> 
  pivot_longer(cols = starts_with("smooth"),
               names_to = "smooth_indicator",
               values_to = "value_smooth") |> 
  ggplot(aes(x = track_release)) +
  geom_jitter(aes(y = danceability,
                  group = seq_along(track_release)),
              width = 20, 
              height = 0.001, 
              alpha = 0.3,
              size = 3,
              color = "#54483e") +
  geom_jitter(aes(y = acousticness,
                  group = seq_along(track_release)),
              width = 20, 
              height = 0.001, 
              alpha = 0.3,
              size = 3,
              color = "#b8396b") +
  ggtext::geom_richtext(data = taylor_albums, 
             mapping = aes(x = album_release,
                           y = 0,
                           label = album_name),
             col = "black",
             angle = 90, 
            hjust = "left",
            alpha = 0.8) +
  geom_line(aes(y = value_smooth,
                col = smooth_indicator),
            alpha = 0.7,
            lwd = 2) +
  theme_minimal() +
  
  scale_x_continuous(breaks = taylor_albums$album_release,
                     labels = format(taylor_albums$album_release, 
                                     "%b %Y")) +
  
  # Using color palettes for package tayloRswift for her albums
  tayloRswift::scale_color_taylor(palette = "lover",
                                  labels = c("Acousticness",
                                             "Danceability") ) +
  
  labs(x = NULL,
       y = "Spotify App Score for songs",
       color = NULL, 
       title = "Taylor Swift's songs over the years",
       subtitle = "After 2015: Increased Acousticness, reduced danceability",
       caption = "Data: taylor R package (W. Jake Thompson). Animation: Aditya Dahiya #TidyTuesday") +
  
  theme(axis.text.x = element_text(angle = 90,
                                   size = 10),
        panel.grid.minor.x = element_blank(),
        panel.grid.minor.y = element_blank(), 
        legend.position = "bottom",
        title = element_text(hjust = 0.5,
                             size = 20),
        legend.text = element_text(size = 20)) +

  
  transition_reveal(track_release) +
  ease_aes("linear") +
  shadow_mark(past = TRUE)

animate(gganim, 
        height = 600,
        width = 800,
        fps = 10, 
        duration = 10,
        start_pause = 3,
        end_pause = 10)

anim_save("docs/taylor_anim1.gif")

An animated line chart (with scatterplot in background) of Taylor Swift’s songs’ acousticness and danceability. The names of albums and release dates are on the bottom x-axis.

---

Step 4: Exploring further for any interesting correlations

vars = c("danceability", "energy", "loudness",
         "speechiness", "acousticness", "instrumentalness",
         "valence", "tempo", "duration_ms")

# Select some variables to examine relations
taylor_all_songs |> 
  select(all_of(vars)) |> 
  GGally::ggpairs()

Lets re-focus on danceability and acoustics: –

taylor_all_songs |> 
  select(album_name, track_name, track_release,
         danceability, acousticness) |> 
  mutate(era = if_else(track_release < ymd("2015-01-01"),
                       "Earlier Era",
                       "Recent Era")) |> 
  ggplot(aes(y = acousticness,
             x = danceability)) +
  geom_point(aes(group = era,
                 col = era),
             alpha = 0.75) +
  geom_smooth(aes(group = era,
                  col = era),
              se = FALSE) +
  theme_minimal()

Trying Cluster Analysis, with some interesting results!

taylor1 <- taylor_all_songs |> 
  select(album_name, track_name, album_release,
         danceability, acousticness, energy, loudness, 
         speechiness, instrumentalness, tempo) |> 
  drop_na()

taylor_cluster <- taylor1 |>
  select(danceability, acousticness) |> 
  as.matrix() |> 
  kmeans(x = _, centers = 2)

taylor1 |> 
  mutate(cluster = as_factor(taylor_cluster$cluster)) |> 
  mutate(era = if_else(album_release < ymd("2013-01-01"),
                       "Pre-2015",
                       "Post-2015")) |> 
  ggplot(aes(y = acousticness,
             x = danceability)) +
  geom_smooth(se = F,
              method = "lm",
              col = "grey",
              alpha = 0.2,
              lwd = 2) +
  geom_point(aes(col = cluster),
             alpha = 0.5,
             size = 2) +
  facet_wrap(~ fct(era,
                   levels = c("Pre-2015", "Post-2015"))) +
  theme_classic() +
  theme(legend.position = "bottom") +
  # Using color palettes for package tayloRswift
  tayloRswift::scale_color_taylor(
    palette = "lover",
    labels = c("Danceability",
               "Acousticness")) +
  labs(x = "Danceability Score (Spotify)",
       y = "Acousticness Score (Spotify)",
       col = "Songs with higher: ")

Step 5: Lastly, the second animation: –

library(magick)
img <- image_read("docs/taylor_jpeg.jpg") |> 
  image_colorize(opacity = 80,
                 color = "white")
  
anim3 <- taylor1 |> 
  mutate(
    cluster = as_factor(taylor_cluster$cluster),
    album_name = fct_reorder(fct(album_name),
                             .x = album_release)
  ) |> 
  ggplot(aes(y = acousticness,
             x = danceability)) +
  annotation_raster(img,
                    xmin = -Inf, xmax = Inf,
                    ymin = -Inf, ymax = Inf) +
  geom_point(aes(col = cluster),
             alpha = 0.9,
             size = 4,
             shape = 19) +
  geom_text(aes(label = paste0(month(album_release, 
                                     label = TRUE), 
                               ", ", 
                               year(album_release)),
                x = 0.25, y = 0.8
                ),
            hjust = "left",
            size = 5
            ) +
  geom_text(aes(label = album_name,
                x = 0.25, y = 0.95),
            hjust = "left",
            size = 8) +
  facet_wrap(~ album_release) +
  facet_null() +
  scale_color_manual(values = c("#54483e",
                                "#b8396b"),
                     labels = c("Danceability",
                                "Acousticness")) +
  labs(x = "Danceability Score (Spotify)",
       y = "Acousticness Score (Spotify)",
       col = "Songs with higher: ",
       title = "Acoustics and Danceability in Taylor Swift's songs",
       subtitle = "Albums \"folklore\" and \"evermore\" were different from all other albums") +
  theme_minimal() +
  theme(legend.position = "bottom",
        plot.title.position = "plot",
        title = element_text(size = 15),
        legend.text = element_text(size = 15)) +
  transition_states(states = album_release,
                    transition_length = 6,
                    state_length = 2) +
  enter_fade() +
  exit_fade()

animate(anim3,
        height = 350,
        width = 470,
        duration = 20,
        fps = 10,
        end_pause = 3,
        start_pause = 1)

anim_save("docs/taylor_anim2.gif")