Autors:

• Paul Leydier

• Wenceslas Sanchez

A retrouver sur Github

Our goal is to scrap artist lyrics on https://search.azlyrics.com and to analyse them.

import pandas as pd
import numpy as np
import re
import matplotlib.pyplot as plt
import seaborn as sns
import nltk

from random_user_agent.user_agent import UserAgent
from random_user_agent.params import SoftwareName, OperatingSystem
import warnings
import unidecode
import requests
from bs4 import BeautifulSoup
import urllib
import urllib.request
from lxml import html
import sys
from random import random

from nltk.tokenize import TreebankWordTokenizer
from nltk.stem.snowball import EnglishStemmer
from nltk.corpus import stopwords
import jellyfish
import tensorflow as tf
import tensorflow_hub as hub
from sklearn.manifold import TSNE
from sklearn.model_selection import train_test_split

sns.set()

nltk.download('stopwords')

[nltk_data] Downloading package stopwords to
[nltk_data]     C:\Users\paull\AppData\Roaming\nltk_data...
[nltk_data]   Package stopwords is already up-to-date!





True

I. WebScraping

1.1. First step webscraping

We are going to build a dataset, which is going to contain artist names, albums names, musics names, production years, and url to connect to the lyrics webpage.

def artist_to_url(artist):
    """ With an artist name, it generates an URL to `https://www.azlyrics.com`. 
        
    We observed a pattern in the URL generation. If the first caracter in artist name is a letter
    then URL should be :
        https://www.azlyrics.com/{first artist name letter}/{artist name}
    else:
        https://www.azlyrics.com/19/{artist name}
   
    Args:
        -artist: the artist name
    
    Output:
        >>> artist_to_url("Schoolboy Q")
        >>> https://www.azlyrics.com/s/schoolboyq.html'
    
    Raises:
        None
        
    """
    url_base= """https://www.azlyrics.com"""
    name= re.sub(' ','',artist)
    # Remove accents like é or à
    name= unidecode.unidecode(name.lower())
    first_l= list(name)[0]
    # Some artist name start with a number
    # and so it generates an URL otherwise
    if first_l.isalpha():
        url= """{}/{}/{}.html""".format(url_base, first_l, name)
    else:
        url= """{}/19/{}.html""".format(url_base, name)
    return url

#test
artist_to_url("Schoolboy Q")

'https://www.azlyrics.com/s/schoolboyq.html'

def transform_proxy_http(proxy):
    """ From a proxy (with his port), we generate an http connexion adress
        
    Args:
        -proxy
        
    Output:
        >>> transform_proxy_http("91.228.8.162:8080")
        >>> "http://91.228.8.162:8080"
        
    Raises:
        None
        
    """
    return "http://"+proxy

# We wollect a daily proxy list
urll= "https://raw.githubusercontent.com/clarketm/proxy-list/master/proxy-list-raw.txt"
urllib.request.urlretrieve(urll, 'data/proxy_list.txt')

listed_proxy= []
f= open("data/proxy_list.txt", "r")
listed_proxy= f.read().split("\n")

#
# NE PAS EXECUTER SI ON VEUT GAGNER DU TEMPS
good_prox= []
# Pour raison de simpliciter, on cherche que les proxy qui ont un port 8080
pat= re.compile('.:8080$')

proxies_list= [l for l in listed_proxy \
               if l in list(filter(pat.findall, listed_proxy))]

for prox in proxies_list:
    # On cherche 20 proxy de bonnes qualités
    if len(good_prox) <= 20:
        try:
            print(prox)
            
            #On génère un User Agent aléatoire pour chaque proxy
            software_names = [SoftwareName.CHROME.value]
            operating_systems = [OperatingSystem.WINDOWS.value, OperatingSystem.LINUX.value]
            user_agent_rotator = UserAgent(software_names=software_names,
                           operating_systems=operating_systems, limit=100)
            
            proxies= {"http":transform_proxy_http(prox),
                    "https":transform_proxy_http(prox)}

            user_agent = user_agent_rotator.get_random_user_agent()
            headers= {"User-Agent":user_agent}

            r = requests.Session()
            r.headers.update(headers)
            r.proxies.update(proxies)

            # Connexion à la page
            page= r.get("https://www.azlyrics.com/", proxies= proxies, headers= headers)

            # Si la connexion est fructueuse, alors le proxy est stocké
            good_prox.append(prox)
        except:
            # Si je ne peux pas me connecter avec ce proxy, alors je teste le suivant
            print("Not Good")
            continue
    else:
        # Si j'ai 20 bon proxy, j'arrète la sélection
        break
print("Fin")

good_prox[:5]

def album_url_year(artist, proxies_list= None, user_agent= True):
    """ Allows us to generate, for an artist, a dataframe with information about his artistic
        productions.
        
    Starting with a artist name, our function collect albums from that artist, and associate an
    production year, and all music's names.
    The user can either choose to use proxy or not. But if an proxiy list is provided, then the
    computing time will incredibly increase.
    
    Args:
        -artist: artist name
        -proxies_list: default value None. Its a list of proxies the user want to add, to make
            his scraping task easier, but time consuming.
        -user_agent: default value True. The user can choose to generate a random user agent.
        
    Output:
        >>> album_url_year("jinjer")
        >>> 	Artiste	Annee	Album	Url
            0	jinjer	2012	Inhale. Do Not Breathe	https://www.azlyrics.com/lyrics/jinjer/untilth...
            1	jinjer	2012	Inhale. Do Not Breathe	https://www.azlyrics.com/lyrics/jinjer/waltz.html
            2	jinjer	2012	Inhale. Do Not Breathe	https://www.azlyrics.com/lyrics/jinjer/scissor...
            3	jinjer	2012	Inhale. Do Not Breathe	https://www.azlyrics.com/lyrics/jinjer/exposed...
            4	jinjer	2012	Inhale. Do Not Breathe	https://www.azlyrics.com/lyrics/jinjer/mylostc...
            5	jinjer	2014	Cloud Factory     	    https://www.azlyrics.com/lyrics/jinjer/outland...
            6	jinjer	2014	Cloud Factory    	    https://www.azlyrics.com/lyrics/jinjer/aplusor...
    
    Raises:
        -warning if we can't built our dataframev (check the list lengths)
        -warning if we can't get the artist (artist not in the website, wrong URL pattern, too
            much error with proxy using)
        -warning if the proxy can't help us to connect
        
    """
    
    year, album_name, url_song= [], [], []
    
    if user_agent:
        software_names = [SoftwareName.CHROME.value]
        operating_systems = [OperatingSystem.WINDOWS.value, OperatingSystem.LINUX.value]
        user_agent_rotator = UserAgent(software_names= software_names,
                                   operating_systems= operating_systems, limit= 100)
        
        user_agent = user_agent_rotator.get_random_user_agent()
        headers= {"User-Agent":user_agent}
        
        r = requests.Session()
        r.headers.update(headers)
    
    else:
        r = requests.Session()
        headers= None
        
    if proxies_list:
        
        random_proxy= sorted(proxies_list, key=lambda x: random())
        i = 0
        for prox in random_proxy:
            if i < 5: #s if too many failed
                i += 1
                try:

                    proxies= {"http":transform_proxy_http(prox),
                            "https":transform_proxy_http(prox)}

                    r.proxies.update(proxies)

                    page= r.get(artist_to_url(artist), proxies= proxies, headers= headers)

                    soup= BeautifulSoup(page.text, 'html.parser')
                    html_page= soup.find('div', id= 'listAlbum')

                    html_data= str(html_page).split('<div class="album"')[1:][:-1]
                    for ht in html_data:
                    # we definie regex pattern to get album name from html page, with years, and
                        # lyrics URL
                        pattern_yr= re.compile('</b> \((\d{4})\)</div>|</b> \((\d{4})\)<br/>')
                        pattern_alb= re.compile('<b>"([\s\S]*)"</b>')
                        pattern_url= re.compile('href="..([\s\S]*).html" target=')

                        year.append(pattern_yr.search(ht).group(1))
                        album_name.append(pattern_alb.search(ht).group(1))
                        
                        url_list= ht.split('\n')
                        url_base= "https://www.azlyrics.com"
                        url_song.append([url_base + pattern_url.search(href).group(1) + ".html" \
                                for href in url_list \
                                if pattern_url.search(href) != None])
                    try:
                        df= pd.DataFrame(
                        {
                            "Artiste": [artist]*len(year),
                            "Annee": year,
                            "Album": album_name,
                            "Url": url_song
                        })
                    except:
                        warnings.warn("Attention, il y a un problème dans la construction du DataFrame")
                        return None
                    
                    # the dataframe we generated contain a list of url
                    # we want 1 line per url
                    df_not_listed= pd.DataFrame({
                      col:np.repeat(df[col].values, df["Url"].str.len())
                      for col in df.columns.drop("Url")
                    }
                    ).assign(**{
                        "Url":np.concatenate(df["Url"].values)
                    })[df.columns]

                    return df_not_listed
                except:
                    warnings.warn("Attention, le proxy {} n'a pas permis de vous connecter à \
                                  la page souhaitée".format(prox))
                    continue
            else:
                break
                
        warnings.warn("Attention, l'artiste {} n'a pas été trouvé".format(artist))
        return None
    
    else:
        
        page= r.get(artist_to_url(artist))
        soup= BeautifulSoup(page.text, 'html.parser')
        html_page= soup.find('div', id= 'listAlbum')
        
        try: # try to connect
            html_data= str(html_page).split('<div class="album"')[1:][:-1]
            for ht in html_data:
                # we definie regex pattern to get album name from html page, with years, and
                # lyrics URL                
                pattern_yr= re.compile('</b> \((\d{4})\)</div>|</b> \((\d{4})\)<br/>')
                pattern_alb= re.compile('<b>"([\s\S]*)"</b>')
                pattern_url= re.compile('href="..([\s\S]*).html" target=')

                year.append(pattern_yr.search(ht).group(1))
                album_name.append(pattern_alb.search(ht).group(1))
                
                url_list= ht.split('\n')
                url_base= "https://www.azlyrics.com"
                url_song.append([url_base + pattern_url.search(href).group(1) + ".html" \
                        for href in url_list \
                        if pattern_url.search(href) != None])
            try: # built the end dataframe
                df= pd.DataFrame(
                {
                    "Artiste": [artist]*len(year),
                    "Annee": year,
                    "Album": album_name,
                    "Url": url_song
                })
                
            except: 
                warnings.warn("Attention, il y a un problème dans la construction du DataFrame")
                return None

            # the dataframe we generated contain a list of url
            # we want 1 line per url
            df_not_listed= pd.DataFrame({
              col:np.repeat(df[col].values, df["Url"].str.len())
              for col in df.columns.drop("Url")
            }
            ).assign(**{
                "Url":np.concatenate(df["Url"].values)
            })[df.columns]

            return df_not_listed
        
        except:
            warnings.warn("Attention, l'artiste {} n'a pas été trouvé".format(artist))
            return None

album_url_year("jinjer", user_agent= False)

artiste_name_style = {
    "Slipknot":"rock",
    "DREAMERS":"rock",
    "Bring Me The Horizon":"rock",
    "Motionless in White":"rock",
    "Falling In Reverse":"rock",
    "Eminem":"rap",
    "Justin Bieber":"pop",
    "Post Malone":"rap",
    "Billie Eilish":"pop",
    "Mac Miller":"rap",
}

artist_name= [name for name in artiste_name_style.keys()]

artist_random= sorted(artist_name, key=lambda x: random())
df= pd.DataFrame()
for art in artist_random:
    print("\n"+art)
    df= pd.concat([album_url_year(art), df], ignore_index= True)

print("\nFin")

Falling In Reverse

Billie Eilish

Bring Me The Horizon

Post Malone

Motionless in White

DREAMERS

Justin Bieber

Slipknot

Eminem

Mac Miller

Fin

#ajout colonne Style à partir du dictionnaire
df["Style"]= df["Artiste"].map(artiste_name_style)
df.head()

def verification(df):
    print("Vérifier si des valeurs sont None")
    for col in df.columns:
        nul= (df[col].isnull()).sum()
        sh= df.shape[0]
        print("\n")
        print(col)
        print(nul)
        print("Soit: {}% de valeurs None".format((nul/sh)*100))
    pass

verification(df)

Vérifier si des valeurs sont None


Artiste
0
Soit: 0.0% de valeurs None


Annee
0
Soit: 0.0% de valeurs None


Album
0
Soit: 0.0% de valeurs None


Url
0
Soit: 0.0% de valeurs None


Style
0
Soit: 0.0% de valeurs None

df.shape

(1064, 5)

# in case, we saved the actual dataset
export_df= df.to_csv("data/dataset_url_lyrics.csv", index= False, header= True)

1.2. Get lyrics from URL

df= pd.read_csv('data/dataset_url_lyrics.csv')
print(df.shape)
df.head()

(1064, 5)

def pretransformation_lyric(url_son, proxies_list= None, user_agent= True):
    """ Get url web page content
    
    Args:
        -url_son: url which contains the lyrics 
        -proxies_list: default value None. Its a list of proxies the user want to add, to make
            his scraping task easier, but time consuming.
        -user_agent: default value True. The user can choose to generate a random user agent.
    
    Output: 
        >>> pretransformation_lyric("https://www.azlyrics.com/lyrics/jinjer/ape.html")
        >>> 'The beginning of all beginnings The era of nonexistence When the Earth wasnt 
            the Earth When our planet was a fireball It was raining heavily for a thousand 
            years And the fireball became an ocean That cooking water the primal soup A natural 
            lifegiving potion  This substance strove to find a way To create another form of 
            miracle  I brought a myriad atoms and turn them into your flesh Regenerated 
            through destruction and chaos While the young world was sculpting itself No one 
            could notice how you stroke a different pose  Voracious ape who walks upright 
            Who deserts life to satisfy his appetite Dont you remember you were the size 
            of a bean Now what a shame to wear the name of a human being  Oh here you are 
            Vertically standing He who calls himself intelligent But has not much for understanding
            Oh here you go Mr Knowitall I made a mistake when I created you long time ago  
            Oh here you go Oh here you go Oh here you go  Just the mention of your name Would 
            make the whole world start to shake Who gave you right to save your life By taking 
            life from others By taking life away  By taking life away By taking life away  
            Bloodthirsty ape who walks upright What shall I do to humble your appetite Bloodthirsty
            ape who walks upright What shall I do to humble your appetite   '
    
    Raises:
        -warning if the proxy can't help us to connect
        
            
    """
    if user_agent: # user agent
        software_names = [SoftwareName.CHROME.value]
        operating_systems = [OperatingSystem.WINDOWS.value, OperatingSystem.LINUX.value]
        user_agent_rotator = UserAgent(software_names= software_names,
                                   operating_systems= operating_systems, limit= 100)
        
        user_agent = user_agent_rotator.get_random_user_agent()
        headers= {"User-Agent":user_agent}
        
        r = requests.Session()
        r.headers.update(headers)
    
    else:
        r = requests.Session()
        headers= None
        
    if proxies_list:
        random_proxy= sorted(proxies_list, key=lambda x: random())
        for prox in random_proxy:
            try:
                proxies= {"http":transform_proxy_http(prox),
                        "https":transform_proxy_http(prox)}
                
                r.proxies.update(proxies)
                page= r.get(url_son, proxies= proxies, headers= headers)

                soup= BeautifulSoup(page.text, 'html.parser')

                lyric= str(page.content)
                pattern= re.compile('(?:Sorry about that. -->)([\s\S]*)(?:-- MxM)')
                try:
                    res= pattern.search(lyric).group(1)
                except: # if the website catch us, we return to the website hobby, so the .group(1)
                        # might not work
                    warnings.warn("Attention, vous êtes arrivé à la page d'acceuil du site")
                    continue # select another proxy
                
                # encoding to get the \n into html page
                res= res.encode('utf8').decode("unicode_escape")
                # delete some html tag
                banword= ['br', 'div', 'brbr']
                # tokenize our text
                tokenizer= TreebankWordTokenizer()
                # we split tokenized elements
                tokenz= [','.join(tokenizer.tokenize(mot)) for mot in res.split()]

                tokenz= [mot.replace(",", "").replace("<br>", "") for mot in tokenz]
                # we remove useless space, and the spescial sign
                tokenz= [re.sub('[^\w\s]', ' ', mot) for mot in tokenz]
                tokenz= [mot.replace(' ','') for mot in tokenz]
                # apply our banword list
                text_clean= ''
                for mot in tokenz:
                    if mot not in banword:
                        text_clean += mot + ' '

                return text_clean

            except:
                warnings.warn("Attention, le proxy {} n'a pas permis de vous connecter à \
                                  la page souhaitée".format(prox))
                continue
    else:
        page= r.get(url_son, headers= headers)
        soup= BeautifulSoup(page.text, 'html.parser')

        lyric= str(page.content)
        pattern= re.compile('(?:Sorry about that. -->)([\s\S]*)(?:-- MxM)')
        
        try:
            res= pattern.search(lyric).group(1)
        except: # if the website catch us, we return to the website hobby, so the .group(1)
                # might not work
            warnings.warn("Attention, vous êtes arrivé à la page d'acceuil du site")
            warnings.warn("Vérifier que l'adresse utilisée n'ait pas été fichée")
            return None
        # encoding to get the \n into html page
        res= res.encode('utf8').decode("unicode_escape")
        # delete some html tag
        banword= ['br', 'div', 'brbr']
        # tokenize our text
        tokenizer= TreebankWordTokenizer()
        # we split tokenized elements
        tokenz= [','.join(tokenizer.tokenize(mot)) for mot in res.split()]

        tokenz= [mot.replace(",", "").replace("<br>", "") for mot in tokenz]
        # we remove useless space, and the spescial sign
        tokenz= [re.sub('[^\w\s]', ' ', mot) for mot in tokenz]
        tokenz= [mot.replace(' ','') for mot in tokenz]
        # apply our banword list
        text_clean= ''
        for mot in tokenz:
            if mot not in banword:
                text_clean += mot + ' '

        return text_clean

print(df.shape)
df = df.sample(frac=1).reset_index(drop=True)
export_df= df.to_csv("random_dataset_url_lyrics.csv", index= False, header= True)
df.head()

(1064, 5)

df["Artiste"].unique()

array(['Eminem', 'DREAMERS', 'Mac Miller', 'Motionless in White',
       'Slipknot', 'Bring Me The Horizon', 'Justin Bieber',
       'Billie Eilish', 'Falling In Reverse', 'Post Malone'], dtype=object)

# We can separate our dataset, and apply the transformation function directly on those sets
# because with proxies, the requests are really time consuming. In case of problems,
# we prefer to make saves

# df_1= df.iloc[:200,:]
# df_2= df.iloc[200:400,:]
# df_3= df.iloc[400:600,:]
# df_4= df.iloc[600:800,:]
# df_5= df.iloc[800:1000,:]
# df_6= df.iloc[1000:df.shape[0],:]

# df_1["test"]= df_1.apply(lambda row: pretransformation_lyric(row[3])
#                      ,axis= 1)

# df["test"]= df.apply(lambda row: pretransformation_lyric(row[3])
#                      ,axis= 1)

II. Lyrics Analysis

# Let's import and merge the csv files generated during the data collection
df= pd.DataFrame()
for i in range(6):
    try: # if file exists
        df= pd.concat([df, pd.read_csv("data/0{}_dataset_lyrics.csv".format(i+1))])
    except:
        #au cas ou , si mon fichier n'existe pas alors j'ajoute un df vide
        df= pd.concat([df, pd.DataFrame()])
        
df= df.rename({"test":"Lyrics"}, axis=1)
df.drop(columns='Url', inplace=True)
df.loc[df['Style'] == 'Rock\\Alternatif', 'Style'] = 'Rock/Alternatif' # Fixing a small issue with the dataset
df.dropna(inplace=True)

df.shape

(1020, 5)

df.head()

As we can see, we have successfully scrapped the lyrics of over 1 thousand songs. We also extracted interesting meta data which are going to enhance our analysis:

• Artist
• Year
• Album name
• Style

1. How many words?

# Splitting the lyrics into arrays of words
df['words'] = df.apply(
    lambda row: np.char.lower(np.array(re.findall(r'\w+', row['Lyrics']))),
    axis=1
)

# Erase the stop words
stop_words = np.array(set(stopwords.words('english')))
df['words'] = df['words'].apply(
    lambda word_array: word_array[np.logical_not(np.isin(word_array, stop_words))]
)

df['words']

0     [im, a, goat, and, for, those, of, yall, who, ...
1     [ooww, ladies, and, gentlemen, the, moment, yo...
2     [this, is, where, we, begin, calling, all, car...
3     [these, ideas, are, nightmares, to, white, par...
4     [alright, lets, go, theres, gonna, be, one, le...
                            ...                        
23    [the, sun, is, silent, in, this, place, the, s...
24    [iintro, mac, milleri, might, as, well, introd...
25    [oh, no, oh, no, oh, they, say, that, hate, ha...
26    [all, the, freaks, coming, out, when, the, sun...
27    [cut, off, my, wings, and, come, lock, me, up,...
Name: words, Length: 1020, dtype: object

1.1. Lyrics length

The first interesting insight that we can get from our dataset is the number of words used.

df['n_words'] = df['words'].apply(len)

plt.figure(figsize=(12,5))
for style in df['Style'].unique():
    words_per_year = df[df['Style'] == style].groupby('Annee')['n_words'].mean()
    sns.lineplot(words_per_year.index, words_per_year.values, label=style)
plt.title('Various styles mean number of words')
plt.ylabel('Mean number of words')
plt.xlabel('Year')
plt.legend()
plt.show()

fig, ax = plt.subplots(4, 1, figsize=(16, 12))
for (i, style) in enumerate(df['Style'].unique()):
    for artist in df.loc[df['Style'] == style, 'Artiste'].unique():
        words_per_year = df[df['Artiste'] == artist].groupby('Annee')['n_words'].mean()
        sns.lineplot(words_per_year.index, words_per_year.values, label=artist, ax=ax[i])
    ax[i].set_title(style)
    ax[i].set_xlabel('')
plt.suptitle('Mean number of words per artist and per style')
plt.legend()
plt.show()

As we can see, most of our music style have a homogeneous number of words around 300, except for the rap genre that is near double, around 600 words per music!

Thus, rap music generally includes more words per song than other music styles.

Nevertheless, there is more heterogeneity in the rap style than in other styles. For example, Eminem produces music with way more lyrics than Post Malone, even though they are both considered rappers.

We should now have a look at the diversity of these words.

1.2. Unique words

Now that we know the sizes of lyrics, let’s have a look at their diversity!

df['n_unique_words'] = df['words'].apply(
    lambda word_list: len(set(word_list))
)

plt.figure(figsize=(12,5))
for style in df['Style'].unique():
    unique_words_per_year = df[df['Style'] == style].groupby('Annee')['n_unique_words'].mean()
    sns.lineplot(unique_words_per_year.index, unique_words_per_year.values, label=style)
plt.title('Various styles mean number of unique words')
plt.ylabel('Mean number of words')
plt.xlabel('Year')
plt.legend()
plt.show()

fig, ax = plt.subplots(4, 1, figsize=(16, 12))
for (i, style) in enumerate(df['Style'].unique()):
    for artist in df.loc[df['Style'] == style, 'Artiste'].unique():
        unique_words_per_year = df[df['Artiste'] == artist].groupby('Annee')['n_unique_words'].mean()
        sns.lineplot(unique_words_per_year.index, unique_words_per_year.values, label=artist, ax=ax[i])
    ax[i].set_title(style)
    ax[i].set_xlabel('')
plt.suptitle('Mean number of unique words per artist and per style')
plt.legend()
plt.show()

We can see here that, once again, the rap style uses more diverse words than the other music styles. This is no surprise, since it also uses more words (even non-unique) per songs, so we are going to look at the uniqueness ratio of the songs.

df['unique_words_ratio'] = df['n_unique_words'] / df['n_words'] # Using highly optimized Numpy broadcasting

plt.figure(figsize=(12,5))
for style in df['Style'].unique():
    unique_words_per_year = df[df['Style'] == style].groupby('Annee')['unique_words_ratio'].mean()
    sns.lineplot(unique_words_per_year.index, unique_words_per_year.values, label=style)
plt.title('Various styles mean ratio of unique words')
plt.ylabel('Mean number of words')
plt.xlabel('Year')
plt.legend()
plt.show()

This graph is interesting! We can see two insights here:

• Even though rap produces a higher quantity of lyrics, the ratio of unique words is not largely higher than other genres.
• There are visible trends in the ratio of unique words:
• Metal diversity of vocabulary decline sinced 1995.
• Pop diversity of vocabulary is rapidly increasing.

fig, ax = plt.subplots(4, 1, figsize=(16, 12))
for (i, style) in enumerate(df['Style'].unique()):
    for artist in df.loc[df['Style'] == style, 'Artiste'].unique():
        unique_words_per_year = df[df['Artiste'] == artist].groupby('Annee')['unique_words_ratio'].mean()
        sns.lineplot(unique_words_per_year.index, unique_words_per_year.values, label=artist, ax=ax[i])
    ax[i].set_title(style)
    ax[i].set_xlabel('')
plt.suptitle('Mean ratio of unique words per artist and per style')
plt.legend()
plt.show()

As we can see, the rapid increase of ratio of unique words comes from Billie Eilish’s work. Is this a general trend in the pop genre, or is this a specificity of our dataset? We would need more data to confirm this.

2. Which words?

Now that we have analyzed the quantities of words in the various lyrics of our dataset, we are going to have a qualitative approach, and understand which words are being used.

2.1. Most commonly used words

First, we are going to have a look at the most commonly used words. Nevertheless, we want to consider the various declinations of a word into one single word, not many. We are thus going to stem our words.

# Stemming the words before counting
stemmer = EnglishStemmer()
df['stem_words'] = df['words'].apply(
    lambda word_list: [stemmer.stem(word) for word in word_list] #TODO: optimize
)

# Counting the used words
all_words = pd.DataFrame({'word': [word for row in df['stem_words'] for word in row], 'count': 1})
all_words = all_words.groupby('word').sum()

plt.figure(figsize=(14,5))
plt.stem(all_words.sort_values('count', ascending=False)[:50].index,
        all_words.sort_values('count', ascending=False)[:50])
plt.xticks(rotation=90)
plt.title('Most comonly used words')
plt.ylabel('Number of use')
plt.show()

C:\dev\anaconda3\lib\site-packages\ipykernel_launcher.py:3: UserWarning: In Matplotlib 3.3 individual lines on a stem plot will be added as a LineCollection instead of individual lines. This significantly improves the performance of a stem plot. To remove this warning and switch to the new behaviour, set the "use_line_collection" keyword argument to True.
  This is separate from the ipykernel package so we can avoid doing imports until

fig, ax = plt.subplots(4, 1, figsize=(14,10))
for (i, style) in enumerate(df['Style'].unique()):
    # Counting the used words
    all_words = pd.DataFrame({'word': [word for row in df.loc[df['Style']==style, 'stem_words'] for word in row], 'count': 1})
    all_words = all_words.groupby('word').sum()
    ax[i].stem(
        all_words.sort_values('count', ascending=False)[:50].index,
        all_words.sort_values('count', ascending=False)[:50]
    )
    ax[i].xaxis.set_tick_params(rotation=90)
    ax[i].set_title(style)
    ax[i].set_ylabel('Number of use')
plt.suptitle('Most commonly used word per style')
plt.tight_layout()
plt.show()

C:\dev\anaconda3\lib\site-packages\ipykernel_launcher.py:8: UserWarning: In Matplotlib 3.3 individual lines on a stem plot will be added as a LineCollection instead of individual lines. This significantly improves the performance of a stem plot. To remove this warning and switch to the new behaviour, set the "use_line_collection" keyword argument to True.

2.2. Vulgarity

We are now going to have a look at the vulgarity of the lyrics of our dataset.

In order to do so, we are going to use a ban word list which is used for facebook moderation.

ban_words = pd.read_csv("data/ban_word.csv", header= None, names= ["nono"])
ban_words = ban_words.drop(index= 0).reset_index(drop= True)
ban_words = np.array(ban_words.values)
print(f'There are {len(ban_words)} banned words in our list.')

There are 1703 banned words in our list.

# Keep only the banned_words
df['banned_words'] = df['words'].apply(
    lambda word_array: word_array[np.isin(word_array, ban_words)]
)

df['n_banned_words'] = df['banned_words'].apply(len)

plt.figure(figsize=(12,5))
for style in df['Style'].unique():
    words_per_year = df[df['Style'] == style].groupby('Annee')['n_banned_words'].mean()
    sns.lineplot(words_per_year.index, words_per_year.values, label=style)
plt.title('Mean number of banned words')
plt.ylabel('Mean number of banned words')
plt.xlabel('Year')
plt.legend()
plt.show()

As we can see, the rap music style is, once again, the highest one on the charts, but this is caused by the cheer number of word in their songs. Let’s have a look at the ratio!

df['banned_words_ratio'] = df['n_banned_words'] / df['n_words'] # Using highly optimized Numpy broadcasting

plt.figure(figsize=(12,5))
for style in df['Style'].unique():
    words_per_year = df[df['Style'] == style].groupby('Annee')['banned_words_ratio'].mean()
    sns.lineplot(words_per_year.index, words_per_year.values, label=style)
plt.title('Mean ratio of banned words')
plt.ylabel('Mean ratio of banned words')
plt.xlabel('Year')
plt.legend()
plt.show()

This graph allows us to draw some conclusions:

• Rap style includes more banned words both proportionnaly and in general.
• Many genres seem to be reducing their use of banned words through the years.
• The pop music style uses really few banned words. This might be because its target audience is wider.

2.3. Vocabulary style

We are now going to try and visualize the vocabulary used by the various lyrics we collected.

# Using the nnlm-en-dim128 TensorFlow embedding model available at
# https://tfhub.dev/google/nnlm-en-dim128/2
# The model is > 450MB, download might take a long time (25 minutes).
# Nevertheless, tensorflow hub caches the model, so the next uses are way faster (2 seconds).
embed = hub.KerasLayer("https://tfhub.dev/google/nnlm-en-dim128/2", input_shape=[], dtype=tf.string)

# Embedding of the Lyrics to a 128 dimensions dense tf.Tensor
embeddings = embed(df['Lyrics'].values)

# t-SNE 2 dimensional projection of the embeddings
projections = TSNE(n_components=2).fit_transform(embeddings)

plt.figure(figsize=(12, 8))
sns.scatterplot(x=projections[:, 0], y=projections[:, 1], hue=df['Style'].values)
plt.title('Lyrics projection by style')

Text(0.5, 1.0, 'Lyrics projection by style')

We can now visualize and interpret the projections of our lyrics.

As we can see, rap lyrics are definitely different than metal lyrics, it is easily visible. On the other hand, metal and rock/alternatif lyrics seem really similar. Finally, pop lyrics seem really diversified and do not visibly differ with other music styles.

fig, ax = plt.subplots(2, 2, figsize=(12, 12), sharex='all', sharey='all')
for (i, style) in enumerate(df['Style'].unique()):
    mask = df['Style'] == style
    sns.scatterplot(
        x=projections[mask, 0],
        y=projections[mask, 1],
        hue=df.loc[mask, 'Artiste'].values,
        ax=ax.flat[i]
    )
    ax.flat[i].set_title(style)
plt.suptitle('Lyrics projection by artist')
plt.tight_layout()
plt.show()

We can dig deeper and visualize artist by artist.

One good application would be to use this dataset to predict music style / artist when given a lyrics.

3. Supervised learning

3.1. Predict music style

We are now going to construct a model based on the embedder used earlier to try and predict the music style of our lyrics.

model = tf.keras.Sequential()
model.add(embed)
model.add(tf.keras.layers.Dense(16, activation='relu'))
model.add(tf.keras.layers.Dense(4, activation='softmax'))
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy', 'AUC'])
model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
keras_layer (KerasLayer)     (None, 128)               124642688 
_________________________________________________________________
dense (Dense)                (None, 16)                2064      
_________________________________________________________________
dense_1 (Dense)              (None, 4)                 68        
=================================================================
Total params: 124,644,820
Trainable params: 2,132
Non-trainable params: 124,642,688
_________________________________________________________________

X = df['Lyrics'].values
y = pd.get_dummies(df['Style']).values

X_train, X_test, y_train, y_test = train_test_split(X, y)

model.fit(X_train, y_train, epochs=100)

Train on 765 samples
Epoch 1/100
765/765 [==============================] - 1s 1ms/sample - loss: 1.1507 - accuracy: 0.5281 - AUC: 0.7556
Epoch 2/100
765/765 [==============================] - 0s 231us/sample - loss: 1.0582 - accuracy: 0.5451 - AUC: 0.7984
Epoch 3/100
765/765 [==============================] - 0s 241us/sample - loss: 0.9818 - accuracy: 0.5647 - AUC: 0.8344
Epoch 4/100
765/765 [==============================] - 0s 253us/sample - loss: 0.9276 - accuracy: 0.6026 - AUC: 0.8554
Epoch 5/100
765/765 [==============================] - 0s 247us/sample - loss: 0.8897 - accuracy: 0.6314 - AUC: 0.8695
Epoch 6/100
765/765 [==============================] - 0s 240us/sample - loss: 0.8539 - accuracy: 0.6588 - AUC: 0.8814
Epoch 7/100
765/765 [==============================] - 0s 253us/sample - loss: 0.8254 - accuracy: 0.6824 - AUC: 0.8934
Epoch 8/100
765/765 [==============================] - 0s 254us/sample - loss: 0.7880 - accuracy: 0.7059 - AUC: 0.9042
Epoch 9/100
765/765 [==============================] - 0s 252us/sample - loss: 0.7653 - accuracy: 0.7085 - AUC: 0.9087
Epoch 10/100
765/765 [==============================] - 0s 255us/sample - loss: 0.7432 - accuracy: 0.7190 - AUC: 0.9169
Epoch 11/100
765/765 [==============================] - 0s 265us/sample - loss: 0.7291 - accuracy: 0.7229 - AUC: 0.9156
Epoch 12/100
765/765 [==============================] - 0s 237us/sample - loss: 0.7092 - accuracy: 0.7333 - AUC: 0.9222
Epoch 13/100
765/765 [==============================] - 0s 259us/sample - loss: 0.6855 - accuracy: 0.7386 - AUC: 0.9269
Epoch 14/100
765/765 [==============================] - 0s 254us/sample - loss: 0.6706 - accuracy: 0.7556 - AUC: 0.9314
Epoch 15/100
765/765 [==============================] - 0s 247us/sample - loss: 0.6569 - accuracy: 0.7634 - AUC: 0.9321
Epoch 16/100
765/765 [==============================] - 0s 253us/sample - loss: 0.6445 - accuracy: 0.7739 - AUC: 0.9367
Epoch 17/100
765/765 [==============================] - 0s 247us/sample - loss: 0.6341 - accuracy: 0.7712 - AUC: 0.9379
Epoch 18/100
765/765 [==============================] - 0s 242us/sample - loss: 0.6226 - accuracy: 0.7712 - AUC: 0.9386
Epoch 19/100
765/765 [==============================] - 0s 242us/sample - loss: 0.6123 - accuracy: 0.7830 - AUC: 0.9424
Epoch 20/100
765/765 [==============================] - 0s 252us/sample - loss: 0.6039 - accuracy: 0.7830 - AUC: 0.9430
Epoch 21/100
765/765 [==============================] - 0s 250us/sample - loss: 0.5952 - accuracy: 0.7791 - AUC: 0.9454
Epoch 22/100
765/765 [==============================] - 0s 240us/sample - loss: 0.5861 - accuracy: 0.7856 - AUC: 0.9461
Epoch 23/100
765/765 [==============================] - 0s 259us/sample - loss: 0.5810 - accuracy: 0.7922 - AUC: 0.9469
Epoch 24/100
765/765 [==============================] - 0s 293us/sample - loss: 0.5756 - accuracy: 0.7987 - AUC: 0.9481
Epoch 25/100
765/765 [==============================] - 0s 235us/sample - loss: 0.5638 - accuracy: 0.7908 - AUC: 0.9497
Epoch 26/100
765/765 [==============================] - 0s 251us/sample - loss: 0.5571 - accuracy: 0.8065 - AUC: 0.9518
Epoch 27/100
765/765 [==============================] - 0s 251us/sample - loss: 0.5563 - accuracy: 0.7974 - AUC: 0.9514
Epoch 28/100
765/765 [==============================] - 0s 244us/sample - loss: 0.5513 - accuracy: 0.8000 - AUC: 0.9507
Epoch 29/100
765/765 [==============================] - 0s 284us/sample - loss: 0.5408 - accuracy: 0.8131 - AUC: 0.9541
Epoch 30/100
765/765 [==============================] - 0s 256us/sample - loss: 0.5301 - accuracy: 0.8157 - AUC: 0.9563
Epoch 31/100
765/765 [==============================] - 0s 266us/sample - loss: 0.5262 - accuracy: 0.8170 - AUC: 0.9563
Epoch 32/100
765/765 [==============================] - 0s 234us/sample - loss: 0.5190 - accuracy: 0.8261 - AUC: 0.9575
Epoch 33/100
765/765 [==============================] - 0s 288us/sample - loss: 0.5128 - accuracy: 0.8235 - AUC: 0.9587
Epoch 34/100
765/765 [==============================] - 0s 276us/sample - loss: 0.5147 - accuracy: 0.8209 - AUC: 0.9584
Epoch 35/100
765/765 [==============================] - 0s 238us/sample - loss: 0.5134 - accuracy: 0.8222 - AUC: 0.9579
Epoch 36/100
765/765 [==============================] - 0s 244us/sample - loss: 0.5047 - accuracy: 0.8353 - AUC: 0.9599
Epoch 37/100
765/765 [==============================] - 0s 245us/sample - loss: 0.4970 - accuracy: 0.8301 - AUC: 0.9612
Epoch 38/100
765/765 [==============================] - 0s 240us/sample - loss: 0.4899 - accuracy: 0.8327 - AUC: 0.9621
Epoch 39/100
765/765 [==============================] - 0s 264us/sample - loss: 0.4861 - accuracy: 0.8405 - AUC: 0.9625
Epoch 40/100
765/765 [==============================] - 0s 244us/sample - loss: 0.4799 - accuracy: 0.8392 - AUC: 0.9638
Epoch 41/100
765/765 [==============================] - 0s 250us/sample - loss: 0.4753 - accuracy: 0.8353 - AUC: 0.9643
Epoch 42/100
765/765 [==============================] - 0s 246us/sample - loss: 0.4729 - accuracy: 0.8431 - AUC: 0.9649
Epoch 43/100
765/765 [==============================] - 0s 245us/sample - loss: 0.4683 - accuracy: 0.8458 - AUC: 0.9653
Epoch 44/100
765/765 [==============================] - 0s 249us/sample - loss: 0.4663 - accuracy: 0.8484 - AUC: 0.9661
Epoch 45/100
765/765 [==============================] - 0s 273us/sample - loss: 0.4586 - accuracy: 0.8484 - AUC: 0.9669
Epoch 46/100
765/765 [==============================] - 0s 266us/sample - loss: 0.4557 - accuracy: 0.8497 - AUC: 0.9668
Epoch 47/100
765/765 [==============================] - 0s 234us/sample - loss: 0.4521 - accuracy: 0.8497 - AUC: 0.9675
Epoch 48/100
765/765 [==============================] - 0s 240us/sample - loss: 0.4517 - accuracy: 0.8562 - AUC: 0.9681
Epoch 49/100
765/765 [==============================] - 0s 242us/sample - loss: 0.4446 - accuracy: 0.8510 - AUC: 0.9684
Epoch 50/100
765/765 [==============================] - 0s 240us/sample - loss: 0.4402 - accuracy: 0.8641 - AUC: 0.9695
Epoch 51/100
765/765 [==============================] - 0s 251us/sample - loss: 0.4391 - accuracy: 0.8614 - AUC: 0.9698
Epoch 52/100
765/765 [==============================] - 0s 226us/sample - loss: 0.4353 - accuracy: 0.8458 - AUC: 0.9698
Epoch 53/100
765/765 [==============================] - 0s 249us/sample - loss: 0.4309 - accuracy: 0.8575 - AUC: 0.9706
Epoch 54/100
765/765 [==============================] - 0s 253us/sample - loss: 0.4245 - accuracy: 0.8601 - AUC: 0.9721
Epoch 55/100
765/765 [==============================] - 0s 253us/sample - loss: 0.4227 - accuracy: 0.8575 - AUC: 0.9715
Epoch 56/100
765/765 [==============================] - 0s 245us/sample - loss: 0.4201 - accuracy: 0.8549 - AUC: 0.9726
Epoch 57/100
765/765 [==============================] - 0s 250us/sample - loss: 0.4189 - accuracy: 0.8627 - AUC: 0.9723
Epoch 58/100
765/765 [==============================] - 0s 242us/sample - loss: 0.4134 - accuracy: 0.8627 - AUC: 0.9731
Epoch 59/100
765/765 [==============================] - 0s 239us/sample - loss: 0.4166 - accuracy: 0.8575 - AUC: 0.9726
Epoch 60/100
765/765 [==============================] - 0s 252us/sample - loss: 0.4070 - accuracy: 0.8680 - AUC: 0.9740
Epoch 61/100
765/765 [==============================] - 0s 243us/sample - loss: 0.4006 - accuracy: 0.8654 - AUC: 0.9748
Epoch 62/100
765/765 [==============================] - 0s 252us/sample - loss: 0.3966 - accuracy: 0.8745 - AUC: 0.9756
Epoch 63/100
765/765 [==============================] - 0s 243us/sample - loss: 0.3927 - accuracy: 0.8693 - AUC: 0.9761
Epoch 64/100
765/765 [==============================] - 0s 245us/sample - loss: 0.3903 - accuracy: 0.8719 - AUC: 0.9764
Epoch 65/100
765/765 [==============================] - 0s 248us/sample - loss: 0.3895 - accuracy: 0.8745 - AUC: 0.9765
Epoch 66/100
765/765 [==============================] - 0s 238us/sample - loss: 0.3900 - accuracy: 0.8680 - AUC: 0.9757
Epoch 67/100
765/765 [==============================] - 0s 242us/sample - loss: 0.3855 - accuracy: 0.8706 - AUC: 0.9769
Epoch 68/100
765/765 [==============================] - 0s 262us/sample - loss: 0.3817 - accuracy: 0.8732 - AUC: 0.9777
Epoch 69/100
765/765 [==============================] - 0s 250us/sample - loss: 0.3831 - accuracy: 0.8824 - AUC: 0.9767
Epoch 70/100
765/765 [==============================] - 0s 247us/sample - loss: 0.3682 - accuracy: 0.8850 - AUC: 0.9795
Epoch 71/100
765/765 [==============================] - 0s 236us/sample - loss: 0.3713 - accuracy: 0.8758 - AUC: 0.9787
Epoch 72/100
765/765 [==============================] - 0s 252us/sample - loss: 0.3676 - accuracy: 0.8732 - AUC: 0.9791
Epoch 73/100
765/765 [==============================] - 0s 244us/sample - loss: 0.3616 - accuracy: 0.8850 - AUC: 0.9805
Epoch 74/100
765/765 [==============================] - 0s 245us/sample - loss: 0.3589 - accuracy: 0.8784 - AUC: 0.9802
Epoch 75/100
765/765 [==============================] - 0s 256us/sample - loss: 0.3548 - accuracy: 0.8810 - AUC: 0.9808
Epoch 76/100
765/765 [==============================] - 0s 241us/sample - loss: 0.3490 - accuracy: 0.8915 - AUC: 0.9820
Epoch 77/100
765/765 [==============================] - 0s 252us/sample - loss: 0.3480 - accuracy: 0.8889 - AUC: 0.9820
Epoch 78/100
765/765 [==============================] - 0s 242us/sample - loss: 0.3512 - accuracy: 0.8928 - AUC: 0.9812
Epoch 79/100
765/765 [==============================] - 0s 246us/sample - loss: 0.3434 - accuracy: 0.8954 - AUC: 0.9822
Epoch 80/100
765/765 [==============================] - 0s 251us/sample - loss: 0.3364 - accuracy: 0.8941 - AUC: 0.9833
Epoch 81/100
765/765 [==============================] - 0s 248us/sample - loss: 0.3310 - accuracy: 0.8993 - AUC: 0.9838
Epoch 82/100
765/765 [==============================] - 0s 234us/sample - loss: 0.3364 - accuracy: 0.8928 - AUC: 0.9828
Epoch 83/100
765/765 [==============================] - 0s 243us/sample - loss: 0.3341 - accuracy: 0.8993 - AUC: 0.9834
Epoch 84/100
765/765 [==============================] - 0s 237us/sample - loss: 0.3286 - accuracy: 0.8928 - AUC: 0.9841
Epoch 85/100
765/765 [==============================] - 0s 240us/sample - loss: 0.3256 - accuracy: 0.9085 - AUC: 0.9844
Epoch 86/100
765/765 [==============================] - 0s 245us/sample - loss: 0.3188 - accuracy: 0.9007 - AUC: 0.9854
Epoch 87/100
765/765 [==============================] - 0s 249us/sample - loss: 0.3254 - accuracy: 0.8941 - AUC: 0.9843
Epoch 88/100
765/765 [==============================] - 0s 247us/sample - loss: 0.3139 - accuracy: 0.9046 - AUC: 0.9856
Epoch 89/100
765/765 [==============================] - 0s 241us/sample - loss: 0.3185 - accuracy: 0.8967 - AUC: 0.9848
Epoch 90/100
765/765 [==============================] - 0s 251us/sample - loss: 0.3078 - accuracy: 0.9046 - AUC: 0.9866
Epoch 91/100
765/765 [==============================] - 0s 243us/sample - loss: 0.3066 - accuracy: 0.9046 - AUC: 0.9863
Epoch 92/100
765/765 [==============================] - 0s 249us/sample - loss: 0.3070 - accuracy: 0.9072 - AUC: 0.9866
Epoch 93/100
765/765 [==============================] - 0s 227us/sample - loss: 0.3075 - accuracy: 0.9020 - AUC: 0.9863
Epoch 94/100
765/765 [==============================] - 0s 248us/sample - loss: 0.2981 - accuracy: 0.9033 - AUC: 0.9871
Epoch 95/100
765/765 [==============================] - 0s 253us/sample - loss: 0.2989 - accuracy: 0.9059 - AUC: 0.9869
Epoch 96/100
765/765 [==============================] - 0s 249us/sample - loss: 0.2945 - accuracy: 0.9190 - AUC: 0.9880
Epoch 97/100
765/765 [==============================] - 0s 269us/sample - loss: 0.2926 - accuracy: 0.9124 - AUC: 0.9879
Epoch 98/100
765/765 [==============================] - 0s 239us/sample - loss: 0.2916 - accuracy: 0.9098 - AUC: 0.9879
Epoch 99/100
765/765 [==============================] - 0s 241us/sample - loss: 0.2882 - accuracy: 0.9176 - AUC: 0.9882
Epoch 100/100
765/765 [==============================] - 0s 284us/sample - loss: 0.2869 - accuracy: 0.9124 - AUC: 0.9883





<tensorflow.python.keras.callbacks.History at 0x20a81c27908>

test_loss, test_acc, test_auc = model.evaluate(X_test, y_test, verbose=2)

255/255 - 0s - loss: 0.6734 - accuracy: 0.7725 - AUC: 0.9330

As we can see, our model which extends the nnlm model is capable, with a small dataset of only a thourand lyrics, of differencing several music styles.

It achieves an accuracy of 79% and a ROC AUC of 0.94!

3.2. Predict Artist

We are now going to try and predict the artist of a lyrics witht the same model.

# Construct the model
model = tf.keras.Sequential()
model.add(embed)
model.add(tf.keras.layers.Dense(16, activation='relu'))
model.add(tf.keras.layers.Dense(len(df['Artiste'].unique()), activation='softmax'))
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy', 'AUC'])

# Prepare the dataset
X = df['Lyrics'].values
y = pd.get_dummies(df['Artiste']).values
X_train, X_test, y_train, y_test = train_test_split(X, y)

# Training
model.fit(X_train, y_train, epochs=100)

Train on 765 samples
Epoch 1/100
765/765 [==============================] - 1s 1ms/sample - loss: 2.2359 - accuracy: 0.1634 - AUC: 0.6277
Epoch 2/100
765/765 [==============================] - 0s 253us/sample - loss: 2.0559 - accuracy: 0.1817 - AUC: 0.7198
Epoch 3/100
765/765 [==============================] - 0s 263us/sample - loss: 1.9833 - accuracy: 0.2170 - AUC: 0.7446
Epoch 4/100
765/765 [==============================] - 0s 253us/sample - loss: 1.9222 - accuracy: 0.2654 - AUC: 0.7651
Epoch 5/100
765/765 [==============================] - 0s 264us/sample - loss: 1.8691 - accuracy: 0.3033 - AUC: 0.7831
Epoch 6/100
765/765 [==============================] - 0s 261us/sample - loss: 1.8190 - accuracy: 0.3216 - AUC: 0.7984
Epoch 7/100
765/765 [==============================] - 0s 275us/sample - loss: 1.7750 - accuracy: 0.3438 - AUC: 0.8107
Epoch 8/100
765/765 [==============================] - 0s 258us/sample - loss: 1.7343 - accuracy: 0.3830 - AUC: 0.8231
Epoch 9/100
765/765 [==============================] - 0s 268us/sample - loss: 1.6958 - accuracy: 0.3987 - AUC: 0.8331
Epoch 10/100
765/765 [==============================] - 0s 257us/sample - loss: 1.6539 - accuracy: 0.4327 - AUC: 0.8441
Epoch 11/100
765/765 [==============================] - 0s 249us/sample - loss: 1.6128 - accuracy: 0.4549 - AUC: 0.8515
Epoch 12/100
765/765 [==============================] - 0s 264us/sample - loss: 1.5732 - accuracy: 0.4824 - AUC: 0.8603
Epoch 13/100
765/765 [==============================] - 0s 257us/sample - loss: 1.5342 - accuracy: 0.5007 - AUC: 0.8706
Epoch 14/100
765/765 [==============================] - 0s 270us/sample - loss: 1.4996 - accuracy: 0.5098 - AUC: 0.8734
Epoch 15/100
765/765 [==============================] - 0s 257us/sample - loss: 1.4646 - accuracy: 0.5359 - AUC: 0.8843
Epoch 16/100
765/765 [==============================] - 0s 301us/sample - loss: 1.4368 - accuracy: 0.5490 - AUC: 0.8879
Epoch 17/100
765/765 [==============================] - 0s 283us/sample - loss: 1.4029 - accuracy: 0.5451 - AUC: 0.8937
Epoch 18/100
765/765 [==============================] - 0s 305us/sample - loss: 1.3746 - accuracy: 0.5608 - AUC: 0.8974
Epoch 19/100
765/765 [==============================] - 0s 271us/sample - loss: 1.3443 - accuracy: 0.5647 - AUC: 0.9044
Epoch 20/100
765/765 [==============================] - 0s 287us/sample - loss: 1.3173 - accuracy: 0.5856 - AUC: 0.9081
Epoch 21/100
765/765 [==============================] - 0s 240us/sample - loss: 1.2931 - accuracy: 0.5922 - AUC: 0.9123
Epoch 22/100
765/765 [==============================] - 0s 256us/sample - loss: 1.2699 - accuracy: 0.5974 - AUC: 0.9158
Epoch 23/100
765/765 [==============================] - 0s 252us/sample - loss: 1.2455 - accuracy: 0.5843 - AUC: 0.9179
Epoch 24/100
765/765 [==============================] - 0s 254us/sample - loss: 1.2231 - accuracy: 0.6131 - AUC: 0.9234
Epoch 25/100
765/765 [==============================] - 0s 265us/sample - loss: 1.2052 - accuracy: 0.6078 - AUC: 0.9239
Epoch 26/100
765/765 [==============================] - 0s 246us/sample - loss: 1.1823 - accuracy: 0.6248 - AUC: 0.9294
Epoch 27/100
765/765 [==============================] - 0s 261us/sample - loss: 1.1654 - accuracy: 0.6196 - AUC: 0.9305
Epoch 28/100
765/765 [==============================] - 0s 290us/sample - loss: 1.1485 - accuracy: 0.6275 - AUC: 0.9320
Epoch 29/100
765/765 [==============================] - 0s 242us/sample - loss: 1.1319 - accuracy: 0.6261 - AUC: 0.9355
Epoch 30/100
765/765 [==============================] - 0s 259us/sample - loss: 1.1158 - accuracy: 0.6340 - AUC: 0.9364
Epoch 31/100
765/765 [==============================] - 0s 267us/sample - loss: 1.0941 - accuracy: 0.6405 - AUC: 0.9408
Epoch 32/100
765/765 [==============================] - 0s 248us/sample - loss: 1.0800 - accuracy: 0.6366 - AUC: 0.9413
Epoch 33/100
765/765 [==============================] - 0s 250us/sample - loss: 1.0603 - accuracy: 0.6497 - AUC: 0.9437
Epoch 34/100
765/765 [==============================] - 0s 251us/sample - loss: 1.0476 - accuracy: 0.6431 - AUC: 0.9454
Epoch 35/100
765/765 [==============================] - 0s 258us/sample - loss: 1.0312 - accuracy: 0.6497 - AUC: 0.9467
Epoch 36/100
765/765 [==============================] - 0s 275us/sample - loss: 1.0147 - accuracy: 0.6706 - AUC: 0.9491
Epoch 37/100
765/765 [==============================] - 0s 278us/sample - loss: 1.0025 - accuracy: 0.6654 - AUC: 0.9497
Epoch 38/100
765/765 [==============================] - 0s 251us/sample - loss: 0.9903 - accuracy: 0.6732 - AUC: 0.9514
Epoch 39/100
765/765 [==============================] - 0s 268us/sample - loss: 0.9773 - accuracy: 0.6850 - AUC: 0.9533
Epoch 40/100
765/765 [==============================] - 0s 272us/sample - loss: 0.9648 - accuracy: 0.6837 - AUC: 0.9535
Epoch 41/100
765/765 [==============================] - 0s 319us/sample - loss: 0.9556 - accuracy: 0.6771 - AUC: 0.9558
Epoch 42/100
765/765 [==============================] - 0s 314us/sample - loss: 0.9477 - accuracy: 0.6824 - AUC: 0.9559
Epoch 43/100
765/765 [==============================] - 0s 284us/sample - loss: 0.9364 - accuracy: 0.6915 - AUC: 0.9565
Epoch 44/100
765/765 [==============================] - 0s 265us/sample - loss: 0.9242 - accuracy: 0.6876 - AUC: 0.9582
Epoch 45/100
765/765 [==============================] - 0s 258us/sample - loss: 0.9155 - accuracy: 0.6902 - AUC: 0.9587
Epoch 46/100
765/765 [==============================] - 0s 252us/sample - loss: 0.9045 - accuracy: 0.6941 - AUC: 0.9603
Epoch 47/100
765/765 [==============================] - 0s 291us/sample - loss: 0.8938 - accuracy: 0.6967 - AUC: 0.9610
Epoch 48/100
765/765 [==============================] - 0s 260us/sample - loss: 0.8828 - accuracy: 0.7072 - AUC: 0.9621
Epoch 49/100
765/765 [==============================] - 0s 319us/sample - loss: 0.8711 - accuracy: 0.7059 - AUC: 0.9630
Epoch 50/100
765/765 [==============================] - 0s 257us/sample - loss: 0.8649 - accuracy: 0.7150 - AUC: 0.9635
Epoch 51/100
765/765 [==============================] - 0s 262us/sample - loss: 0.8579 - accuracy: 0.7137 - AUC: 0.9644
Epoch 52/100
765/765 [==============================] - 0s 263us/sample - loss: 0.8512 - accuracy: 0.7072 - AUC: 0.9644
Epoch 53/100
765/765 [==============================] - 0s 253us/sample - loss: 0.8388 - accuracy: 0.7281 - AUC: 0.9664
Epoch 54/100
765/765 [==============================] - 0s 258us/sample - loss: 0.8334 - accuracy: 0.7255 - AUC: 0.9663
Epoch 55/100
765/765 [==============================] - 0s 281us/sample - loss: 0.8276 - accuracy: 0.7373 - AUC: 0.9664
Epoch 56/100
765/765 [==============================] - 0s 256us/sample - loss: 0.8178 - accuracy: 0.7320 - AUC: 0.9681
Epoch 57/100
765/765 [==============================] - 0s 272us/sample - loss: 0.8115 - accuracy: 0.7307 - AUC: 0.9681
Epoch 58/100
765/765 [==============================] - 0s 257us/sample - loss: 0.8032 - accuracy: 0.7346 - AUC: 0.9690
Epoch 59/100
765/765 [==============================] - 0s 261us/sample - loss: 0.7942 - accuracy: 0.7399 - AUC: 0.9698
Epoch 60/100
765/765 [==============================] - 0s 272us/sample - loss: 0.7883 - accuracy: 0.7516 - AUC: 0.9703
Epoch 61/100
765/765 [==============================] - 0s 260us/sample - loss: 0.7896 - accuracy: 0.7294 - AUC: 0.9698
Epoch 62/100
765/765 [==============================] - 0s 265us/sample - loss: 0.7725 - accuracy: 0.7425 - AUC: 0.9719
Epoch 63/100
765/765 [==============================] - 0s 265us/sample - loss: 0.7711 - accuracy: 0.7477 - AUC: 0.9716
Epoch 64/100
765/765 [==============================] - 0s 257us/sample - loss: 0.7601 - accuracy: 0.7582 - AUC: 0.9725
Epoch 65/100
765/765 [==============================] - 0s 262us/sample - loss: 0.7531 - accuracy: 0.7529 - AUC: 0.9731
Epoch 66/100
765/765 [==============================] - 0s 263us/sample - loss: 0.7476 - accuracy: 0.7569 - AUC: 0.9735
Epoch 67/100
765/765 [==============================] - 0s 254us/sample - loss: 0.7448 - accuracy: 0.7660 - AUC: 0.9734
Epoch 68/100
765/765 [==============================] - 0s 261us/sample - loss: 0.7386 - accuracy: 0.7516 - AUC: 0.9740
Epoch 69/100
765/765 [==============================] - 0s 267us/sample - loss: 0.7347 - accuracy: 0.7660 - AUC: 0.9745
Epoch 70/100
765/765 [==============================] - 0s 250us/sample - loss: 0.7237 - accuracy: 0.7634 - AUC: 0.9751
Epoch 71/100
765/765 [==============================] - 0s 251us/sample - loss: 0.7168 - accuracy: 0.7686 - AUC: 0.9761
Epoch 72/100
765/765 [==============================] - 0s 262us/sample - loss: 0.7093 - accuracy: 0.7686 - AUC: 0.9767
Epoch 73/100
765/765 [==============================] - 0s 247us/sample - loss: 0.7063 - accuracy: 0.7673 - AUC: 0.9767
Epoch 74/100
765/765 [==============================] - 0s 267us/sample - loss: 0.7002 - accuracy: 0.7791 - AUC: 0.9774
Epoch 75/100
765/765 [==============================] - 0s 255us/sample - loss: 0.6970 - accuracy: 0.7739 - AUC: 0.9772
Epoch 76/100
765/765 [==============================] - 0s 259us/sample - loss: 0.6924 - accuracy: 0.7778 - AUC: 0.9776
Epoch 77/100
765/765 [==============================] - 0s 252us/sample - loss: 0.6855 - accuracy: 0.7817 - AUC: 0.9782
Epoch 78/100
765/765 [==============================] - 0s 257us/sample - loss: 0.6833 - accuracy: 0.7843 - AUC: 0.9781
Epoch 79/100
765/765 [==============================] - 0s 258us/sample - loss: 0.6727 - accuracy: 0.7948 - AUC: 0.9790
Epoch 80/100
765/765 [==============================] - 0s 251us/sample - loss: 0.6695 - accuracy: 0.7895 - AUC: 0.9796
Epoch 81/100
765/765 [==============================] - 0s 272us/sample - loss: 0.6610 - accuracy: 0.7987 - AUC: 0.9799
Epoch 82/100
765/765 [==============================] - 0s 250us/sample - loss: 0.6626 - accuracy: 0.7935 - AUC: 0.9797
Epoch 83/100
765/765 [==============================] - 0s 276us/sample - loss: 0.6572 - accuracy: 0.7987 - AUC: 0.9799
Epoch 84/100
765/765 [==============================] - 0s 265us/sample - loss: 0.6543 - accuracy: 0.7922 - AUC: 0.9804
Epoch 85/100
765/765 [==============================] - 0s 236us/sample - loss: 0.6402 - accuracy: 0.8039 - AUC: 0.9815
Epoch 86/100
765/765 [==============================] - 0s 266us/sample - loss: 0.6386 - accuracy: 0.8052 - AUC: 0.9815
Epoch 87/100
765/765 [==============================] - 0s 268us/sample - loss: 0.6330 - accuracy: 0.7948 - AUC: 0.9818
Epoch 88/100
765/765 [==============================] - 0s 272us/sample - loss: 0.6318 - accuracy: 0.8026 - AUC: 0.9818
Epoch 89/100
765/765 [==============================] - 0s 288us/sample - loss: 0.6275 - accuracy: 0.8052 - AUC: 0.9823
Epoch 90/100
765/765 [==============================] - 0s 264us/sample - loss: 0.6195 - accuracy: 0.8065 - AUC: 0.9826
Epoch 91/100
765/765 [==============================] - 0s 257us/sample - loss: 0.6208 - accuracy: 0.8092 - AUC: 0.9823
Epoch 92/100
765/765 [==============================] - 0s 280us/sample - loss: 0.6174 - accuracy: 0.8078 - AUC: 0.9829
Epoch 93/100
765/765 [==============================] - 0s 245us/sample - loss: 0.6080 - accuracy: 0.8105 - AUC: 0.9833
Epoch 94/100
765/765 [==============================] - 0s 264us/sample - loss: 0.6072 - accuracy: 0.8092 - AUC: 0.9833
Epoch 95/100
765/765 [==============================] - 0s 289us/sample - loss: 0.6017 - accuracy: 0.8157 - AUC: 0.9839
Epoch 96/100
765/765 [==============================] - 0s 250us/sample - loss: 0.5995 - accuracy: 0.8131 - AUC: 0.9840
Epoch 97/100
765/765 [==============================] - 0s 262us/sample - loss: 0.5971 - accuracy: 0.8078 - AUC: 0.9836
Epoch 98/100
765/765 [==============================] - 0s 279us/sample - loss: 0.5879 - accuracy: 0.8183 - AUC: 0.9846
Epoch 99/100
765/765 [==============================] - 0s 319us/sample - loss: 0.5896 - accuracy: 0.8105 - AUC: 0.9843
Epoch 100/100
765/765 [==============================] - 0s 339us/sample - loss: 0.5837 - accuracy: 0.8209 - AUC: 0.9848





<tensorflow.python.keras.callbacks.History at 0x20a85f98048>

# Evaluating the results
test_loss, test_acc, test_auc = model.evaluate(X_test, y_test, verbose=2)

255/255 - 0s - loss: 1.0791 - accuracy: 0.6157 - AUC: 0.9323

Our model achieves good results here too, even though less good than on the music style prediction, which is probably easier to modelise.

The next step to improve our models is to collect more data and restart the training.