Slay: Predicting song artist based on lyrics

Application exercise
Modified

May 2, 2024

library(tidyverse)
library(tidymodels)
library(stringr)
library(textrecipes)
library(textdata)
library(discrim)
library(themis)
library(vip)

# set seed for randomization
set.seed(123)

theme_set(theme_minimal(base_size = 13))

Import data

lyrics <- read_csv(file = "data/beyonce-swift-lyrics.csv") |>
  mutate(artist = factor(artist))
lyrics

Split the data into analysis/assessment/test sets

Demonstration:

  • Split the data into training/test sets with 75% allocated for training
  • Split the training set into 10 cross-validation folds
# split into training/testing
set.seed(123)
lyrics_split <- initial_split(data = lyrics, strata = artist, prop = 0.75)

lyrics_train <- training(lyrics_split)
lyrics_test <- testing(lyrics_split)

# create cross-validation folds
lyrics_folds <- vfold_cv(data = lyrics_train, strata = artist)

Estimate the null model for a baseline comparison

Demonstration: Estimate a null model to determine an appropriate baseline for evaluating a model’s performance.

null_spec <- null_model() |>
  set_engine("parsnip") |>
  set_mode("classification")

null_spec |>
  fit_resamples(
    # pick something as the predictor - doesn't really matter
    artist ~ danceability,
    resamples = lyrics_folds
  ) |>
  collect_metrics()

Fit a random forest model

Define the feature engineering recipe

Demonstration:

  • Define a feature engineering recipe to predict the song’s artist as a function of the lyrics + audio features
  • Exclude the ID variables from the recipe
  • Tokenize the song lyrics
  • Remove stop words
  • Only keep the 500 most frequently appearing tokens
  • Calculate tf-idf scores for the remaining tokens

    • This will generate one column for every token. Each column will have the standardized name tfidf_lyrics_* where * is the specific token. Instead we would prefer the column names simply be *. You can remove the tfidf_lyrics_ prefix using

      # Simplify these names
step_rename_at(starts_with("tfidf_lyrics_"),
  fn = \(x) str_replace_all(
    string = x,
    pattern = "tfidf_lyrics_",
    replacement = ""
  )
)

    • This does cause a conflict between the energy audio feature and the token energy. Before removing the "tfidf_lyrics_" prefix, we will add a prefix to the audio features to avoid this conflict.

      # Simplify these names
step_rename_at(
  all_predictors(), -starts_with("tfidf_lyrics_"),
  fn = \(x) str_glue("af_{x}")
)
  • Downsample the observations so there are an equal number of songs by Beyoncé and Taylor Swift in the analysis set
# define preprocessing recipe
rf_rec <- recipe(artist ~ ., data = lyrics_train) |>
  TODO
rf_rec

Fit the model

Demonstration:

  • Define a random forest model grown with 1000 trees using the ranger engine.
  • Define a workflow using the feature engineering recipe and random forest model specification. Fit the workflow using the cross-validation folds.
    • Use control = control_resamples(save_pred = TRUE) to save the assessment set predictions. We need these to assess the model’s performance.
    • Use control = control_resamples(save_workflow = TRUE) to save the workflow object. We need this later on if we want to fit a single model using the same workflow and the entire training set.
# define the model specification
rf_spec <- rand_forest(trees = 1000) |>
  set_mode("classification") |>
  # calculate feature importance metrics using the ranger engine
  set_engine("ranger", importance = "permutation")

# define the workflow
rf_wf <- workflow() |>
  add_recipe(rf_rec) |>
  add_model(rf_spec)

# fit the model to each of the cross-validation folds
rf_cv <- rf_wf |>
  fit_resamples(
    resamples = lyrics_folds,
    control = control_resamples(save_pred = TRUE, save_workflow = TRUE, verbose = TRUE)
  )

Evaluate model performance

Demonstration:

  • Calculate the model’s accuracy and ROC AUC. How did it perform?
  • Draw the ROC curve for each validation fold
  • Generate the resampled confusion matrix for the model and draw it using a heatmap. How does the model perform predicting Beyoncé songs relative to Taylor Swift songs?
# extract metrics and predictions
rf_cv_metrics <- collect_metrics(rf_cv)
rf_cv_predictions <- collect_predictions(rf_cv)

# how well did the model perform?
rf_cv_metrics

# roc curve
rf_cv_predictions |>
  group_by(id) |>
  roc_curve(truth = artist, .pred_Beyoncé) |>
  autoplot()

# confusion matrix
conf_mat_resampled(x = rf_cv, tidy = FALSE) |>
  autoplot(type = "heatmap")

Add response here.

Penalized regression

Define the feature engineering recipe

Demonstration:

  • Define a feature engineering recipe to predict the song’s artist as a function of the lyrics + audio features
  • Exclude the ID variables from the recipe
  • Tokenize the song lyrics
  • Calculate all possible 1-grams, 2-grams, 3-grams, 4-grams, and 5-grams
  • Remove stop words
  • Only keep the 2000 most frequently appearing tokens
  • Calculate tf-idf scores for the remaining tokens
  • Rename audio feature and tf-idf as before
  • Apply required steps for penalized regression models
    • Convert the explicit variable to a factor
    • Convert nominal predictors to dummy variables
    • Get rid of zero-variance predictors
    • Normalize all predictors to mean of 0 and variance of 1
  • Downsample the observations so there are an equal number of songs by Beyoncé and Taylor Swift in the analysis set
glmnet_rec <- recipe(artist ~ ., data = lyrics_train) |>
  TODO
glmnet_rec

Tune the penalized regression model

Demonstration:

  • Define the penalized regression model specification, including tuning placeholders for penalty and mixture
  • Create the workflow object
  • Define a tuning grid with every combination of:
    • penalty = 10^seq(-6, -1, length.out = 20)
    • mixture = c(0, 0.2, 0.4, 0.6, 0.8, 1)
  • Tune the model using the cross-validation folds
  • Evaluate the tuning procedure and identify the best performing models based on ROC AUC
# define the penalized regression model specification
glmnet_spec <- logistic_reg(penalty = tune(), mixture = tune()) |>
  set_mode("classification") |>
  set_engine("glmnet")

# define the new workflow
glmnet_wf <- workflow() |>
  add_recipe(glmnet_rec) |>
  add_model(glmnet_spec)

# create the tuning grid
glmnet_grid <- expand_grid(
  penalty = 10^seq(-6, -1, length.out = 20),
  mixture = c(0, 0.2, 0.4, 0.6, 0.8, 1)
)

# tune over the model hyperparameters
glmnet_tune <- tune_grid(
  object = glmnet_wf,
  resamples = lyrics_folds,
  grid = glmnet_grid,
  control = control_grid(save_pred = TRUE, save_workflow = TRUE, verbose = TRUE)
)
# evaluate results
collect_metrics(x = glmnet_tune)
autoplot(glmnet_tune)

# identify the five best hyperparameter combinations
show_best(x = glmnet_tune, metric = "roc_auc")

Add response here.

Random forest with word embeddings

Define the feature engineering recipe

Demonstration: Rather than using individual tokens and tf-idf scores, we will use pre-trained word embeddings to represent the lyrics. We will calculate the aggregate embedding for each song by averaging the embeddings of all the words in the song, then tune a random forest model using these embeddings.

Note

Normally in order to import GloVe embeddings you would use the code below:

glove_embed <- embedding_glove6b(dimensions = 100)

This downloads the ZIP file containing the embeddings, stores it in a cache folder, and then imports the requested embeddings and dimensions as a data frame. Note that many of the embeddings are stored in ZIP files that are multiple gigabytes in size. Often it is easier to manually download the files and store them in the appropriate location outside of R. See the documentation for embedding_glove*() for more information.

# hacky way to make it work on RStudio Workbench
glove_embed <- read_delim(
    file = "/rstudio-files/glove6b/glove.6B.100d.txt",
    delim = " ",
    quote = "",
    col_names = c(
      "token",
      paste0("d", seq_len(100))
    ),
    col_types = paste0(
      c(
        "c",
        rep("d", 100)
      ),
      collapse = ""
    )
  )

rf_embeds_rec <- recipe(artist ~ ., data = lyrics_train) |>
  # exclude ID variables
  update_role(album_name, track_number, track_name, new_role = "id vars") |>
  step_tokenize(lyrics) |>
  # calculate aggregate embedding for each song
  step_word_embeddings(lyrics, embeddings = glove_embed, aggregation = "mean") |>
  step_downsample(artist)
rf_embeds_rec

Tune the model

Demonstration:

  • Define the penalized regression model specification, including tuning placeholders for mtry and min_n
  • Create the workflow object
  • Tune the model using the cross-validation folds and an automatically generated tuning grid
  • Evaluate the tuning procedure and identify the best performing models based on ROC AUC
# define the model specification
rf_embeds_spec <- rand_forest(trees = 1000, mtry = tune(), min_n = tune()) |>
  set_mode("classification") |>
  # calculate feature importance metrics using the ranger engine
  set_engine("ranger", importance = "permutation")

# define the workflow
rf_embeds_wf <- workflow() |>
  add_recipe(rf_embeds_rec) |>
  add_model(rf_embeds_spec)

# fit the model to each of the cross-validation folds
rf_embeds_cv <- rf_embeds_wf |>
  tune_grid(
    resamples = lyrics_folds,
    control = control_grid(save_pred = TRUE, save_workflow = TRUE, verbose = TRUE)
  )
# extract metrics
rf_embeds_cv_metrics <- collect_metrics(rf_embeds_cv)

# how well did the model perform?
rf_embeds_cv_metrics
show_best(rf_embeds_cv, metric = "roc_auc")

Add response here.

Fit the best model

Your turn:

  • Select the model + hyperparameter combinations that achieve the highest ROC AUC
  • Fit that model using the best hyperparameters and the full training set. How well does the model perform on the test set?
# select the best model's hyperparameters
best_fit <- fit_best(TODO)

# test set ROC AUC
bind_cols(
  lyrics_test,
  predict(best_fit, new_data = lyrics_test, type = "prob")
) |>
  roc_auc(truth = artist, .pred_Beyoncé)

Add response here.

Variable importance

We can examine the results of each model to evaluate which tokens were the most important in generating artist predictions. Here we use vip to calculate importance.

# extract parsnip model fit
rf_imp <- rf_cv |>
  fit_best() |>
  extract_fit_parsnip() |>
  vi(method = "model")

# clean up the data frame for visualization
rf_imp |>
  # extract 20 most important n-grams
  slice_max(order_by = Importance, n = 20) |>
  mutate(Variable = fct_reorder(.f = Variable, .x = Importance)) |>
  ggplot(mapping = aes(
    x = Importance,
    y = Variable
  )) +
  geom_col() +
  labs(
    y = NULL,
    title = "Most relevant features for predicting whether\na song is by Beyoncé or Taylor Swift",
    subtitle = "Random forest model"
  )

# extract parsnip model fit
glmnet_imp <- glmnet_tune |>
  fit_best() |>
  extract_fit_parsnip() |>
  vi(method = "model", lambda = select_best(x = glmnet_tune, metric = "roc_auc")$penalty)

# clean up the data frame for visualization
glmnet_imp |>
  mutate(
    Sign = case_when(
      Sign == "NEG" ~ "More likely from Beyoncé",
      Sign == "POS" ~ "More likely from Taylor Swift"
    ),
    Importance = abs(Importance)
  ) |>
  # importance must be greater than 0
  filter(Importance > 0) |>
  # keep top 20 features for each artist
  slice_max(n = 20, order_by = Importance, by = Sign) |>
  mutate(Variable = fct_reorder(.f = Variable, .x = Importance)) |>
  ggplot(mapping = aes(
    x = Importance,
    y = Variable,
    fill = Sign
  )) +
  geom_col(show.legend = FALSE) +
  scale_fill_brewer(type = "qual") +
  facet_wrap(facets = vars(Sign), scales = "free_y") +
  labs(
    y = NULL,
    title = "Most relevant features for predicting whether\na song is by Beyoncé or Taylor Swift",
    subtitle = "Penalized regression model"
  )