knitr::include_graphics('Francesco.JPG')

…

INTRODUCTION

This project proposes and seeks to show that song lyrics can be a determing factor of a song’s genre. I do not discount the musical power of harmony, melody, and rhythm to differentiate genre, but it’s not common for musicians, educators and others to classify music limited to a song’s lyrics. I accept this challenge.

BUSINESS CASE

The main question is: Why would anyone or any company care whether lyrics define the song’s genre?

Let’s look at potential uses. Businesses such at Pandora and Spotify have been analyzing and classifying music (including songs with lyrics) for their customers, to provide an inclusive listening experience. But, many other business can benefit from music and lyrics analysis.

Let’s look at songwriters, their management, licensing companies such as ASCAP and BMI, and music libraries such as HD Publishing.

For songwriters, there is a sub-group that are only lyricists, meaning someone else composes the music. Composers look for lyrics to compose to. Lyricists look for composers who are known for specific genres. Many haven’t worked together before. Managers often have knowledge of both sides, and will broker deals that combine their talents. Giving songwriters and composers the ability look for each other by matching lyrics with music could be a further democratization of the music industry. It could remove the role of the manager (and associated fees) to connect songwriters with music composers and bands.

Licensing companies deal with both lyrics and songs as legal assets they manage for artists. While this is not the main part of their business, they would benefit from validating the lyric’s genre compared to what the artist might think it is.

When another artist wants to record someone else’s song, they have to pay license fees to a licensing company such as ASCAP. ASCAP has a website that deals with this from a financial aspect, but not from a search and song selection aspect. Having this capability would enhance their service offering to their clients, would be a competitive differentiator; and potentially increase profits through “premium” chargeable services.

GOAL

The goal of this project is to use data science to analyze a library of popular song lyrics, put the data in a clean and useable form, and build and train a model to classify the words that best characterize the song’s genre. Ultimately, I plan to be able to use song lyrics not in the sample DB to test the model, using lyrics from major artists and others.

METHODOLOGY

I will use the readily available subset of lyrics from The Million Song library. My sample size will be 57,000 songs from multiple genres and artists. From this, I will cut it into training and test data to build and test my model. I will use packages that are designed to be used for Natural Language Processing (NLP) including tidytext, syuzhet(sentiment) and others.

To train the prediction, I will need to give the model the most likely genre for each song. This is known as supervised training. Unfortunately, this is not provided with the freely available data. Due to the huge task to select the genre for each of the 57,000 songs, I will focus on the artist’s best known genre using multiple readily available sources that categorize music. I have done this for 632 artists.

Note: I have combined some genres to make the analysis work better due to similarities in genres.

Hip-Hop-Rap = Hip-Hop + Rap
D-DJ-E-D-E-H = Disco + DJ Beats + Electronica + Dance + EDM + House
others…
Read 57,000 song dataset from an online .csv file https://www.researchgate.net/publication/220723656_The_Million_Song_Dataset
Read my artist-to-genre mapping .csv file
Join them together into one dataframe
Clean the data
Make Ready for Next Steps by adding an id variable and dropping the URL link variable which is not needed for analysis.

# head(raw_dat)
# kable(head(raw_dat))
  #column_spec(5:7, bold = T) %>%
  #row_spec(3:5, bold = T, color = "white", background = "#D7261E")
# table_output(raw_data)

Wrangle Data

Add Sentiment

Compute: Number of Words, Letters, and Average Word Length per Song.

dat = raw_dat %>%
  mutate(sentiment = get_sentiment(text),
         number_of_words = stri_count(text, regex="\\S+"),
         number_of_letters = nchar(text),
         avg_word_length = number_of_letters / number_of_words)

EDA

Take a Peek at the Data

table_output(raw_dat)

artist	song	text	genre	id
ABBA	Ahe’s My Kind Of Girl	Look at he	Pop	1
ABBA	Andante, Andante	Take it ea	Pop	2
ABBA	As Good As New	I’ll never	Pop	3
ABBA	Bang	Making som	Pop	4
ABBA	Bang-A-Boomerang	Making som	Pop	5
ABBA	Burning My Bridges	Well, you	Pop	6
ABBA	Cassandra	Down in th	Pop	7
ABBA	Chiquitita	Chiquitita	Pop	8
ABBA	Crazy World	I was out	Pop	9
ABBA	Crying Over You	I’m waitin	Pop	10

table_output(dat)

artist	song	text	genre	id	sentiment	number_of_words	number_of_letters	avg_word_length
ABBA	Ahe’s My Kind Of Girl	Look at he	Pop	1	3.95	153	761	4.973856
ABBA	Andante, Andante	Take it ea	Pop	2	5.90	260	1434	5.515385
ABBA	As Good As New	I’ll never	Pop	3	4.15	312	1477	4.733974
ABBA	Bang	Making som	Pop	4	7.95	200	1247	6.235000
ABBA	Bang-A-Boomerang	Making som	Pop	5	8.20	198	1263	6.378788
ABBA	Burning My Bridges	Well, you	Pop	6	-2.10	109	570	5.229358
ABBA	Cassandra	Down in th	Pop	7	-7.10	361	2028	5.617729
ABBA	Chiquitita	Chiquitita	Pop	8	-2.50	304	1591	5.233553
ABBA	Crazy World	I was out	Pop	9	-0.45	304	1533	5.042763
ABBA	Crying Over You	I’m waitin	Pop	10	-0.25	123	637	5.178862

Plot Mean Sentiment for Genres

dat %>%
  filter(sentiment != 0) %>%
  group_by(genre) %>%
  summarize(mean_sentiment = mean(sentiment)) %>%
  ggplot(aes(x = fct_reorder(genre, mean_sentiment), y = mean_sentiment, fill=mean_sentiment)) + 
  geom_col() + 
  coord_flip() + 
  labs(x = '', y = 'Sentiment', title = 'Lyric Sentiment by Genre') + 
  theme(legend.position = 'none')

Most genres have a positive sentiment.

The genres with the most positive sentiment are: Religous, R&B, and Blues.
The genres with the most negative sentiment are: Hip-Hop-Rap and Metal (just slightly negative).
Jazz is close to neutral

Sentiment Density Plot

dat %>%
  ggplot(aes(x = sentiment, col = genre, fill = genre)) +
  geom_density(alpha = 0.5) + 
  labs(x = 'Sentiment', y = 'Density', title = 'Distribution of Sentiment by Genre')

The distribution plot gives us a deeper looking into the genres’ sentiment. All genres seem to have a somewhat normal distribution of sentiments (except Hip-Hop-Rap), but have different means and SDs. Hip-Hop-Rap and Metal are the most left skewed (negative).

Programmatically Identify Key Words in Genres

Exceeded R’s object capacity when unnesting all 57,000 songs (millions of words).

Need to break into 4 pieces, unnest into words, and stitch back together

Keep the important words per genre included.

# Break text_dat_raw into 4 parts
text_dat_raw_1 = raw_dat[1:10000,]
text_dat_raw_2 = raw_dat[10001:20000,]
text_dat_raw_3 = raw_dat[20001:30000,]
text_dat_raw_4 = raw_dat[30001:nrow(raw_dat),]
#Unnest all 4 parts and remove stop_words

text_dat_raw_1 = text_dat_raw_1%>%
unnest_tokens(word, text)

text_dat_raw_2 = text_dat_raw_2%>%
unnest_tokens(word, text)

text_dat_raw_3 = text_dat_raw_3%>%
unnest_tokens(word, text)

text_dat_raw_4 = text_dat_raw_4%>%
unnest_tokens(word, text)


# Combine 4 parts back into 1
text_dat_raw = text_dat_raw_1 %>%
  bind_rows(text_dat_raw_2) %>%
  bind_rows(text_dat_raw_3) %>%
  bind_rows(text_dat_raw_4)


all_ids = text_dat_raw %>% filter(!is.na(id)) %>% select(id) %>% unique()
train_ids = sample_frac(all_ids, 0.75)
test_ids = all_ids %>% anti_join(train_ids)

text_dat_raw_train = train_ids %>% left_join(text_dat_raw)

text_dat_grouped = text_dat_raw_train %>%
  group_by(genre, id, word) %>%
  summarize(n = 1) %>%
  group_by(genre, word) %>%
  summarize(n = n()) %>%
  group_by(genre) %>%
  top_n(50) %>%
  ungroup() %>%
  group_by(word) %>%
  mutate(wdcount = n()) %>%
  ungroup() %>%
  filter(wdcount == 1) %>% 
  select(word)

text_dat_spread = text_dat_grouped %>% 
  left_join(text_dat_raw, by = 'word') %>%
  select(id, word) %>%
  distinct(id, word) %>%
  mutate(n = 1) %>%
  spread(key = word, value = n)

text_dat_spread[is.na(text_dat_spread)]=0

Finalize Data

final_dat = dat %>% select(-song) %>% 
  left_join(text_dat_spread, by = 'id') %>%
  select(-text, -artist) %>%
  janitor::clean_names() %>%
  rename(Class = genre) %>%
  mutate(Class = as.factor(Class))
final_dat[is.na(final_dat)] = 0
table_output(final_dat)

Class	id	sentiment	number_of_words	number_of_letters	avg_word_length	aint	as	back	god	ill	let	lord	take	well
Pop	1	3.95	153	761	4.973856	0	0	0	0	0	0	0	0	0
Pop	2	5.90	260	1434	5.515385	0	0	0	0	0	1	0	1	0
Pop	3	4.15	312	1477	4.733974	0	1	0	1	1	0	0	0	0
Pop	4	7.95	200	1247	6.235000	0	1	0	0	1	0	0	1	0
Pop	5	8.20	198	1263	6.378788	0	1	0	0	1	0	0	1	0
Pop	6	-2.10	109	570	5.229358	1	0	0	0	0	0	0	1	1
Pop	7	-7.10	361	2028	5.617729	0	1	0	0	0	0	0	0	0
Pop	8	-2.50	304	1591	5.233553	0	0	0	0	0	1	0	0	0
Pop	9	-0.45	304	1533	5.042763	0	1	1	0	0	0	0	1	0
Pop	10	-0.25	123	637	5.178862	0	0	0	0	0	0	1	0	0

Build Model

Steps:

Split data 75% to Training Data… 25% to Testing Data
Use genre (i.e “Class”) as the dependent variable. Do not use “genre” or “id” as predictors in model training!
Balance the number words per genre by using Up Sampling technique.

training_split = 0.75
smp_size = floor(training_split * nrow(final_dat))
dat_index = sample(seq_len(nrow(final_dat)), size = smp_size)
dat_train = as.data.frame(final_dat[dat_index,])

dat_train = train_ids %>% left_join(final_dat)
dat_test = test_ids %>% left_join(final_dat)

dat_train = upSample(x = dat_train %>% select(-Class),
                     y = dat_train$Class)

dat_train = dat_train %>%
  select(-id)

Training Data

table_output(dat_train)

sentiment	number_of_words	number_of_letters	avg_word_length	aint	as	back	let	name	take	well	yeah	Class
-1.70	134	787	5.873134	0	0	0	0	1	1	0	0	Blues
-0.15	96	506	5.270833	0	0	1	1	0	0	0	1	Blues
0.45	85	492	5.788235	0	0	1	0	0	0	0	0	Blues
1.25	133	716	5.383459	0	1	0	0	0	0	0	0	Blues
3.05	119	681	5.722689	0	1	0	0	0	0	0	0	Blues
1.95	239	1210	5.062761	0	0	0	0	0	0	0	0	Blues
3.00	384	2232	5.812500	0	0	0	0	0	0	0	1	Blues
1.35	162	853	5.265432	0	0	0	0	0	0	0	0	Blues
3.30	208	1088	5.230769	0	1	0	0	0	0	1	1	Blues
-0.20	163	895	5.490798	1	0	0	0	0	1	0	1	Blues

dat_test = dat_test %>%
  select(-id)

Test Data

table_output(dat_test)

Class	sentiment	number_of_words	number_of_letters	avg_word_length	as	back	god	ill	let	take	well	yeah
Pop	0.00	345	1823	5.284058	0	0	1	0	1	0	0	0
Pop	0.65	368	2169	5.894022	1	1	0	0	0	0	0	0
Pop	2.05	237	1279	5.396624	1	0	0	0	0	0	1	0
Pop	1.00	356	1892	5.314607	0	0	0	0	1	0	1	1
Pop	1.00	248	1457	5.875000	1	0	0	0	0	0	0	0
Pop	5.90	206	1082	5.252427	1	1	0	1	0	0	0	0
Pop	5.85	383	2408	6.287206	0	0	0	1	0	1	0	0
Pop	1.85	214	1208	5.644860	0	0	0	0	0	0	0	0
Pop	-1.25	126	729	5.785714	1	1	0	0	0	0	0	0
Pop	1.85	202	1071	5.301980	1	0	0	0	0	0	0	0

Define Random Forest Model

Use Cross Validation across variables to avoid correlations
Don’t include genre (i.e. “Class”)
ranger = Random Forest
Use 50 trees
Use “impurity” method for brancing classification
Run the prediction model
Report the confusion matrix

train_control = trainControl(method = "cv")

model_rf = train(dat_train %>% select(-Class),
            dat_train$Class,
            method = "ranger",
            num.trees = 50,
            importance = "impurity",
            trControl = train_control)

predictions_rf = predict(model_rf, dat_test)
confusionMatrix(predictions_rf, as.factor(dat_test$Class))

## Confusion Matrix and Statistics
## 
##               Reference
## Prediction     Blues Country D-DJ-E-D-E-H Folk Hip-Hop-Rap Jazz Metal  Pop
##   Blues            4      10            2   10           1    1     1   32
##   Country         14      86           10   52           4   17     4  170
##   D-DJ-E-D-E-H     1       3            3    8           1    0     1   18
##   Folk            10      49            6   49           1   14     9  140
##   Hip-Hop-Rap      1       7            1    7         136    1     0   81
##   Jazz             3      14            1   17           2   12     4   63
##   Metal            1       4            0    5           1    2     6   25
##   Pop             69     260           73  305          73   76    59 1542
##   R&B              2       8            3   10           5    2     1   45
##   Religious        3      20            1    9           3    3     0   46
##   Rock            77     246           51  267          24   79    86 1030
##               Reference
## Prediction      R&B Religious Rock
##   Blues           2         0   27
##   Country        15        17  180
##   D-DJ-E-D-E-H    0         1   21
##   Folk           12         8  152
##   Hip-Hop-Rap    17         1   35
##   Jazz            6         4   47
##   Metal           3         0   26
##   Pop           110        38 1083
##   R&B             9         1   37
##   Religious       2       129   34
##   Rock           56        32 1492
## 
## Overall Statistics
##                                         
##                Accuracy : 0.377         
##                  95% CI : (0.367, 0.387)
##     No Information Rate : 0.347         
##     P-Value [Acc > NIR] : 1.004e-09     
##                                         
##                   Kappa : 0.1372        
##                                         
##  Mcnemar's Test P-Value : NA            
## 
## Statistics by Class:
## 
##                      Class: Blues Class: Country Class: D-DJ-E-D-E-H
## Sensitivity             0.0216216       0.121641           0.0198675
## Specificity             0.9904603       0.943130           0.9940325
## Pos Pred Value          0.0444444       0.151142           0.0526316
## Neg Pred Value          0.9801317       0.928050           0.9838128
## Prevalence              0.0201087       0.076848           0.0164130
## Detection Rate          0.0004348       0.009348           0.0003261
## Detection Prevalence    0.0097826       0.061848           0.0061957
## Balanced Accuracy       0.5060410       0.532385           0.5069500
##                      Class: Folk Class: Hip-Hop-Rap Class: Jazz
## Sensitivity             0.066306            0.54183    0.057971
## Specificity             0.952606            0.98313    0.982097
## Pos Pred Value          0.108889            0.47387    0.069364
## Neg Pred Value          0.921143            0.98710    0.978398
## Prevalence              0.080326            0.02728    0.022500
## Detection Rate          0.005326            0.01478    0.001304
## Detection Prevalence    0.048913            0.03120    0.018804
## Balanced Accuracy       0.509456            0.76248    0.520034
##                      Class: Metal Class: Pop Class: R&B Class: Religious
## Sensitivity             0.0350877     0.4831  0.0387931          0.55844
## Specificity             0.9925795     0.6428  0.9872881          0.98651
## Pos Pred Value          0.0821918     0.4181  0.0731707          0.51600
## Neg Pred Value          0.9819218     0.7007  0.9754324          0.98860
## Prevalence              0.0185870     0.3470  0.0252174          0.02511
## Detection Rate          0.0006522     0.1676  0.0009783          0.01402
## Detection Prevalence    0.0079348     0.4009  0.0133696          0.02717
## Balanced Accuracy       0.5138336     0.5629  0.5130406          0.77248
##                      Class: Rock
## Sensitivity               0.4761
## Specificity               0.6789
## Pos Pred Value            0.4337
## Neg Pred Value            0.7149
## Prevalence                0.3407
## Detection Rate            0.1622
## Detection Prevalence      0.3739
## Balanced Accuracy         0.5775

Plot the Importance of Variables

model_rf$finalModel %>%
  # extract variable importance metrics
  ranger::importance() %>%
  # convert to a data frame
  enframe(name = "variable", value = "varimp") %>%
  top_n(n = 20, wt = varimp) %>%
  # plot the metrics
  ggplot(aes(x = fct_reorder(variable, varimp), y = varimp, fill=varimp)) +
  geom_col() +
  coord_flip() +
  labs(x = "Token",
       y = "Variable importance (higher is more important)")

Save the Random Forest Model for later use.

a = dat_train %>% as_tibble()
model_rf_data_structure = a[0,]
timestamp = Sys.time()
saveRDS(model_rf_data_structure, paste0('models/model_rf_data_structure - ', timestamp, '.rds'))
saveRDS(model_rf, file = paste0('models/model_rf - ', timestamp, '.rds'))


saveRDS(model_rf_data_structure, paste0('models/model_rf_data_structure', '.rds'))
saveRDS(model_rf, file = paste0('models/model_rf', '.rds'))

Produce a Word Cloud

set.seed(1234)
word_count = text_dat_raw %>%
  filter(genre == "Rock" | genre == "Hip-Hop-Rap" | genre == "Religious" | genre == "Country" | genre == "Pop") %>%
  anti_join(stop_words) %>%
  count(word, genre, sort = TRUE) %>%
  mutate(genre = as.character(genre)) 

word_count %>%
  
  reshape2::acast(word ~ genre, value.var = "n", fill = 0) %>%
  comparison.cloud(colors = brewer.pal(12,"Dark2"),
                   max.words = 200,
                   title.size = 2,
                   scale = c(0.2,1.5))
title(main = "Word Cloud from Lyrics")

Conclusion

To a resonable degree, lyrics alone can be used to predict a song’s genre. I acheived a 42% prediction accuracy across the ? genres in my statistical prediction model. This is better than pure guessing (which would be closer to ?%). And some genres can be predicted better than others.

Here are the results of the top 11 genres:

Religious – 77%
Hip-Hop-Rap – 76%
Rock – 58%
Pop – 56%
Country - 53%
Jazz & Metal - 52%
Folks - 51%
Blues – 50%
others…

Engineered variables including number of words, number of letters, and sentiment led the list of importance in the prediction model. While down, like, ain’t, baby, and love were the most important words.

Most genres have a positive sentiment with the exception of Hip-Hop-Rap, and Metal. More precise analysis could further differentiate sentiment into categories such foul language, cultural meanings, and other categories. For example Raggae music commonly uses words such as ra and wit.

Random Forest is a good prediction model for this Natural Language Processing project. And further tuning might get the accuracy higher, but considerable compute resources are needed to run this model. I was able to run an optional Neural Network model that achieved 48% accuracy, but the results weren’t stable enough for me to chose it for my conclusion. And, I did use a Random Forest pre-processing method for the Neural Networks model, showing that they can work together.

Dealing with Big Data in this project was a challenge, but taught me a great deal about several techniques to help with both the data preparation and model running stages. With Big Data there is no substitute for more compute power and memory. But, even with those, limitations in the R language required techniques to work around them.

Combining my lyrics prediction analysis with existing music analysis would likely produce an even higher prediction accuracy. This will be a personal project for me after I leave this certification training.

On a personal note, Data Science has sparked an interest in me to seek work in this field that combines it with my previous experience.

Music by Numbers

Capstone Project

Francesco Bonifazi

9/05/2019