Vectorisation des noms et des descriptions d'articles (textes en anglais)

Notebook basé sur:
Getting Started with Text Vectorization
How to Vectorize Text in DataFrames for NLP Tasks — 3 Simple Techniques
Multimodal deep learning to predict movie genres

import pandas as pd
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.preprocessing import MultiLabelBinarizer
import texthero as hero
from texthero import preprocessing
import en_core_web_sm
import time
import nltk
nltk.download('punkt')
nltk.download('wordnet')
from nltk import word_tokenize
from nltk.stem.wordnet import WordNetLemmatizer
import dask.dataframe as dd
from dask.diagnostics import ProgressBar

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

0. Cleaning data

We use the Texthero library. To apply the default text cleaning script run hero.clean(pandas.Series). It runs the following seven functions by default when using clean():

• fillna(s) Replace not assigned values with empty spaces.
• lowercase(s) Lowercase all text.
• remove_digits() Remove all blocks of digits.
• remove_punctuation() Remove all string.punctuation (!"#$%&'()*+,-./:;<=>?@[]^_`{|}~).
• remove_diacritics() Remove all accents from strings.
• remove_stopwords() Remove all stop words.
• remove_whitespace() Remove all white space between words.

Word lemmatization using NLTK. The goal of lemmatization is to reduce inflectional forms and sometimes derivationally related forms of a word to a common base form. For example, ‘Finds’ can be replaced with just ‘Find’.

Texthero documentation

def contains_digits(w):
    """
    Test de la présence de chiffres dans un mot

    Args:
        w: mot à analyser

    Returns:
        booléen indiquant la présence d'un chiffre
    """
    if not w.isalpha():
        return sum([w.find(str(i))!=-1 for i in range(10) if w.find(str(i-1))==-1])>0
    return False

def clean_text(df):
    """
    Nettoyage des chaines de caractères avant vectorisation

    Args:
        df: dataframe contenant les chaînes de caractères à nettoyer

    Returns:
        dataframe avec les chaînes de caractères nettoyés
    """
    start=time.time()
    lemmatizer = WordNetLemmatizer()
    df_clean = df.apply(lambda s: hero.clean(s))
    df_clean = df_clean.apply(lambda s: s.apply(lambda t: [lemmatizer.lemmatize(w) for w in word_tokenize(t) if not contains_digits(w)]).str.join(' '))
    df_clean = df_clean.apply(lambda s: s.apply(lambda t: ''.join(filter(lambda c: str.isalpha(c)|str.isspace(c), t))))
    df_clean = df_clean.apply(lambda s: hero.clean(s,[preprocessing.remove_whitespace]))
    print(time.strftime('Temps d\'exécution du nettoyage: %Hh %Mmin %Ss', time.gmtime(time.time()-start)))
    return df_clean

try:
    train_clean = pd.read_csv("data/vectors/train_clean.csv")
except:
    train = pd.read_csv("data/traduction/train_trad.csv")
    train_clean = clean_text(train)
    train_clean.to_csv("data/vectors/train_clean.csv",index=False)
train_clean["item_description"]= (train_clean.item_name.fillna('')+" "+train_clean.item_caption.fillna('')).str.strip()
print("Nb NaN:",train_clean.isna().sum(),sep="\n")
train_clean.head()

Nb NaN:
item_name               4
item_caption        24184
item_description        0
dtype: int64

	item_name	item_caption	item_description
0	sankyo aluminum shade beam standing type one s...	item manufacturer sankyo aluminum size width x...	sankyo aluminum shade beam standing type one s...
1	sale sale fashion coordination thick sash belt...	increased presence thick sash belt us horse le...	sale sale fashion coordination thick sash belt...
2	geta paulownia made japan woman tone nose widt...	item paulownia clog yukata half width obi yuka...	geta paulownia made japan woman tone nose widt...
3	limited time yen coupon issuance shoe box widt...	product description louver shoe box width sing...	limited time yen coupon issuance shoe box widt...
4	post mailbox mailbox post multi family housing...	post apartment variable push lock collective m...	post mailbox mailbox post multi family housing...

try:
    test_clean = pd.read_csv("data/vectors/test_clean.csv")
except:
    test = pd.read_csv("data/traduction/test_trad.csv")
    test_clean = clean_text(test)
    test_clean.to_csv("data/vectors/test_clean.csv",index=False)
test_clean["item_description"]= (test_clean.item_name.fillna('')+" "+test_clean.item_caption.fillna('')).str.strip()
print("Nb NaN:",test_clean.isna().sum(),sep="\n")
test_clean.head()

Nb NaN:
item_name              3
item_caption        4286
item_description       0
dtype: int64

	item_name	item_caption	item_description
0	miraie f miraie forte au au smartphone case sm...	precaution depending arrival time material cas...	miraie f miraie forte au au smartphone case sm...
1	xperia premium xperia premium docomo docomo sm...	precaution depending arrival time material cas...	xperia premium xperia premium docomo docomo sm...
2	mo mono mono docomo docomo notebook type smart...	product feature seamless full scale design cal...	mo mono mono docomo docomo notebook type smart...
3	xperia xz notebook type case beach hawaii expe...	compatible model xperia xz xperia sony compati...	xperia xz notebook type case beach hawaii expe...
4	used comme ca du mode skirt bomb toss long len...	used comme ca du mode skirt bomb toss long len...	used comme ca du mode skirt bomb toss long len...

1. Binary Term Frequency

Binary Term Frequency captures presence (1) or absence (0) of term in document. Under TfidfVectorizer, we set binary parameter equal to true so that it can show just presence (1) or absence (0) and norm parameter equal to false.
TfidfVectorizer documentation

start = time.time()
tv =  TfidfVectorizer(binary=True, norm=False, 
        use_idf=False, smooth_idf=False, 
        lowercase=True, stop_words="english",
        min_df=100,max_df=1.0,
        max_features=None,ngram_range=(1,1))

df = pd.DataFrame(tv.fit_transform(train_clean.item_name.fillna('').to_list()).toarray(),columns=tv.get_feature_names())
df = df.astype('int')
df = df.apply(lambda x: x.to_list(),axis=1)
print(time.strftime('Temps d\'exécution: %Hh %Mmin %Ss', time.gmtime(time.time()-start)))
df

Temps d'exécution: 00h 00min 23s

	ab	absorbent	absorbing	absorption	ac	accent	accepted	accessory	accordion	ace	...	zero	zeta	zett	zip	zipper	zippy	zirconia	zone	zoom	zori
0	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
1	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
2	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
3	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
4	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
212115	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
212116	0	0	0	0	0	0	0	1	0	0	...	0	0	0	0	0	0	0	0	0	0
212117	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
212118	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0
212119	0	0	0	0	0	0	0	0	0	0	...	0	0	0	0	0	0	0	0	0	0

212120 rows × 2845 columns

2. Bag of Words (BoW) Term Frequency

Bag of Words (BoW) Term Frequency captures frequency of term in document. Under TfidfVectorizer, we set binary parameter equal to false so that it can show the actual frequency of the term and norm parameter equal to none.

start = time.time()
tv =  TfidfVectorizer(binary=False, norm=None, 
        use_idf=False, smooth_idf=False, 
        lowercase=True, stop_words="english",
        min_df=100,max_df=1.0,
        max_features=None,ngram_range=(1,1))

df = pd.DataFrame(tv.fit_transform(train_clean.item_name.to_list()).toarray(), columns = tv.get_feature_names())
print(time.strftime('Temps d\'exécution: %Hh %Mmin %Ss', time.gmtime(time.time()-start)))
df

Temps d'exécution: 00h 00min 04s

	ab	absorbent	absorbing	absorption	ac	accent	accepted	accessory	accordion	ace	...	zero	zeta	zett	zip	zipper	zippy	zirconia	zone	zoom	zori
0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
1	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
2	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
3	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
4	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
212115	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212116	0.0	0.0	0.0	0.0	0.0	0.0	0.0	1.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212117	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212118	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212119	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0

212120 rows × 2845 columns

3. (L1) Normalized Term Frequency

(L1) Normalized Term Frequency captures normalized BoW term frequency in document. Under TfidfVectorizer, we set binary parameter equal to false so that it can show the actual frequency of the term and norm parameter equal to l1.

start = time.time()
tv =  TfidfVectorizer(binary=False, norm='l1', 
        use_idf=False, smooth_idf=False, 
        lowercase=True, stop_words="english",
        min_df=100,max_df=1.0,
        max_features=None,ngram_range=(1,1))

df = pd.DataFrame(tv.fit_transform(train_clean.item_name.to_list()).toarray(), columns = tv.get_feature_names())
print(time.strftime('Temps d\'exécution: %Hh %Mmin %Ss', time.gmtime(time.time()-start)))
df

Temps d'exécution: 00h 00min 05s

	ab	absorbent	absorbing	absorption	ac	accent	accepted	accessory	accordion	ace	...	zero	zeta	zett	zip	zipper	zippy	zirconia	zone	zoom	zori
0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
1	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
2	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
3	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
4	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
212115	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212116	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.032258	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212117	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212118	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212119	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0

212120 rows × 2845 columns

4.(L2) Normalized TF-IDF

(L2) Normalized TFIDF (Term Frequency–Inverse Document Frequency) captures normalized TFIDF in document. The below is the formula for how to compute the TFIDF.
Under TfidfVectorizer, we set binary parameter equal to false so that it can show the actual frequency of the term and norm parameter equal to l2.

start = time.time()
tv =  TfidfVectorizer(binary=False, norm='l2', 
        use_idf=True, smooth_idf=True, 
        lowercase=True, stop_words="english",
        min_df=100,max_df=0.8,
        max_features=None,ngram_range=(1,1))

df = pd.DataFrame(tv.fit_transform(train_clean.item_name.to_list()).toarray(), columns = tv.get_feature_names())
print(time.strftime('Temps d\'exécution: %Hh %Mmin %Ss', time.gmtime(time.time()-start)))
df

Temps d'exécution: 00h 00min 10s

	ab	absorbent	absorbing	absorption	ac	accent	accepted	accessory	accordion	ace	...	zero	zeta	zett	zip	zipper	zippy	zirconia	zone	zoom	zori
0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
1	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
2	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
3	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
4	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
212115	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212116	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.114688	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212117	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212118	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
212119	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.000000	0.0	0.0	...	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0

212120 rows × 2845 columns

5. Word2Vec

Word2Vec provides embedded representation of words. Word2Vec starts with one representation of all words in the corpus and train a NN (with 1 hidden layer) on a very large corpus of data. Python’s spacy package provides pre-trained models we can use to see how w2v works.

start = time.time()
nlp = en_core_web_sm.load()
vectors = train.item_name.apply(lambda x: nlp(x).vector)
print(time.strftime('Temps d\'exécution: %Hh %Mmin %Ss', time.gmtime(time.time()-start)))
vectors

Temps d'exécution: 00h 25min 37s

0         [0.8806767, 0.5209342, 0.22395355, -0.26502186...
1         [1.4003121, 0.7583619, 0.3834568, -0.5043833, ...
2         [0.96584576, 0.47329023, 0.05513578, -0.224261...
3         [1.426747, 0.83528644, 0.34153345, -0.3079996,...
4         [0.6590709, 0.53607327, -0.050336536, -0.42528...
                                ...                        
212115    [1.2364701, 0.37272543, 0.14906088, -0.2349578...
212116    [0.934301, 0.5912743, 0.29895964, -0.28099176,...
212117    [1.7777156, 0.73792523, 0.06944884, -0.5251924...
212118    [0.13715902, 0.27609712, 0.23340255, -0.275412...
212119    [1.4836605, 0.617889, -0.21495572, -0.4181979,...
Name: item_name, Length: 212120, dtype: object

6. Choix de la vectorisation

def matrix_to_serie(m,Type=None):
    """
    Conversion d'une sparse matrice en série

    Args:
        m: matrice à convertir
        Type: type attendus en sortie

    Returns:
        série issue de la conversion
    """
    df=pd.DataFrame(m)
    try:
        s = df.astype(Type).apply(lambda row: row.to_list(),axis=1)
    except:
        s = df.apply(lambda row: row.to_list(),axis=1)
    return s

def vectorize(df,var,Type="binary",NaN='',min_freq=100,max_freq=1.0,size_max=None):
    """
    Vectorisation des chaînes de caractères

    Args:
        df: dataframe à traiter
        var: liste des variable(s) à vectoriser
        Type: modalité de calcul des vecteurs
        NaN: format des données manquantes
        min_freq: féquence minimale des mots utilisés dans le vocabulaire
        max_freq: féquence maximale des mots utilisés dans le vocabulaire
        size_max: taille maximale des vecteurs

    Returns:
        dataframe avec les colonnes var vectorisées
    """
    start = time.time()
    vectors = df.copy()
    if Type=="binary":
        tv = TfidfVectorizer(binary=True, norm=None, 
            use_idf=False, smooth_idf=False, 
            stop_words="english", max_features=size_max,
            min_df=min_freq,max_df=max_freq)
        for col in var:
            vectors[col+"_"+Type] = matrix_to_serie(tv.fit_transform(vectors[col].fillna(NaN).to_list()).toarray(),'int')
    elif Type=='bow':
        tv =  TfidfVectorizer(norm=None, 
                use_idf=False, smooth_idf=False, 
                stop_words="english", max_features=size_max,
                min_df=min_freq,max_df=max_feq)
    elif Type=='l1':
        tv =  TfidfVectorizer(norm='l1', 
                use_idf=False, smooth_idf=False, 
                stop_words="english", max_features=size_max,
                min_df=min_freq,max_df=max_freq)
    elif Type=='l2':
        tv =  TfidfVectorizer(stop_words="english",max_features=size_max,
                min_df=min_freq,max_df=max_freq)
    elif Type=='w2v':
        nlp = en_core_web_sm.load()
        for col in var:
            vectors[col+"_"+Type] = vectors[col].fillna(NaN).apply(lambda x: nlp(x).vector)
    if Type not in ['w2v','binary']:
        for col in var:
            vectors[col+"_"+Type] = matrix_to_serie(tv.fit_transform(vectors[col].fillna(NaN).to_list()).toarray())
    for col in vectors.columns:
        if Type in col:
            print("Taille des vecteurs de {}:".format(col),len(vectors.loc[0,col]))
    print(time.strftime('Temps d\'exécution ({} vectorization): %Hh %Mmin %Ss'.format(Type.capitalize()), time.gmtime(time.time()-start)))
    return vectors

train_clean2 = vectorize(train_clean,["item_description"],Type="w2v")
train_clean2.drop('item_description',axis='columns').to_csv("data/vectors/train_clean.csv",index=False)
train_clean2.head()

Taille des vecteurs de item_description_w2v: 96
Temps d'exécution (W2v vectorization): 01h 04min 16s

	item_name	item_caption	item_description	item_description_w2v
0	sankyo aluminum shade beam standing type one s...	item manufacturer sankyo aluminum size width x...	sankyo aluminum shade beam standing type one s...	[1.0962551, 0.41255596, 0.079079345, -0.069799...
1	sale sale fashion coordination thick sash belt...	increased presence thick sash belt us horse le...	sale sale fashion coordination thick sash belt...	[1.1558363, 0.6076791, -0.017795345, -0.047616...
2	geta paulownia made japan woman tone nose widt...	item paulownia clog yukata half width obi yuka...	geta paulownia made japan woman tone nose widt...	[0.80632937, 0.38488936, 0.07236661, -0.058630...
3	limited time yen coupon issuance shoe box widt...	product description louver shoe box width sing...	limited time yen coupon issuance shoe box widt...	[1.2116121, 0.62559044, 0.1324889, -0.06670587...
4	post mailbox mailbox post multi family housing...	post apartment variable push lock collective m...	post mailbox mailbox post multi family housing...	[0.9028818, 0.563046, 0.076704845, -0.05939380...

test_clean2 = vectorize(test_clean,["item_description"],Type="w2v")
test_clean2.drop('item_description',axis='columns').to_csv("data/vectors/test_clean.csv",index=False)
test_clean2.head()

Taille des vecteurs de item_description_w2v: 96
Temps d'exécution (W2v vectorization): 00h 14min 01s

	item_name	item_caption	item_description	item_description_w2v
0	miraie f miraie forte au au smartphone case sm...	precaution depending arrival time material cas...	miraie f miraie forte au au smartphone case sm...	[0.7333057, 0.17599767, 0.24099274, -0.0315121...
1	xperia premium xperia premium docomo docomo sm...	precaution depending arrival time material cas...	xperia premium xperia premium docomo docomo sm...	[0.75016546, 0.19974689, 0.2601111, -0.0260416...
2	mo mono mono docomo docomo notebook type smart...	product feature seamless full scale design cal...	mo mono mono docomo docomo notebook type smart...	[0.7519761, 0.2823713, 0.17091945, -0.05869386...
3	xperia xz notebook type case beach hawaii expe...	compatible model xperia xz xperia sony compati...	xperia xz notebook type case beach hawaii expe...	[1.1669095, 0.5273239, 0.2156976, -0.08899468,...
4	used comme ca du mode skirt bomb toss long len...	used comme ca du mode skirt bomb toss long len...	used comme ca du mode skirt bomb toss long len...	[0.8762712, 0.45833683, 0.18252678, -0.0383652...

7. Encodage des labels (y_train)

train_labels = pd.read_csv("data/y_train.csv")
train_labels.color_tags = train_labels.color_tags.str.strip("[]'").str.split("', '")
mlb = MultiLabelBinarizer()
mlb.fit(train_labels.color_tags.tolist())
mlb.classes_

# Output - array(['Beige', 'Black', 'Blue', 'Brown', 'Burgundy', 'Gold', 'Green',
#      'Grey', 'Khaki', 'Multiple Colors', 'Navy', 'Orange', 'Pink',
#      'Purple', 'Red', 'Silver', 'Transparent', 'White', 'Yellow'],
#      dtype=object)

train_labels = mlb.transform(train_labels['color_tags'].to_list())
train_labels

array([[0, 1, 0, ..., 0, 0, 0],
       [0, 1, 0, ..., 0, 0, 0],
       [0, 1, 0, ..., 0, 1, 0],
       ...,
       [1, 0, 0, ..., 0, 0, 0],
       [0, 0, 0, ..., 0, 1, 0],
       [0, 0, 1, ..., 0, 0, 0]])