I am using the label encoder to convert categorical data into numeric values.
How does LabelEncoder handle missing values?
from sklearn.preprocessing import LabelEncoder
import pandas as pd
import numpy as np
a = pd.DataFrame(['A','B','C',np.nan,'D','A'])
le = LabelEncoder()
le.fit_transform(a)
Output:
array([1, 2, 3, 0, 4, 1])
For the above example, label encoder changed NaN values to a category. How would I know which category represents missing values?
Don't use LabelEncoder with missing values. I don't know which version of scikit-learn you're using, but in 0.17.1 your code raises TypeError: unorderable types: str() > float().
As you can see in the source it uses numpy.unique against the data to encode, which raises TypeError if missing values are found. If you want to encode missing values, first change its type to a string:
a[pd.isnull(a)] = 'NaN'
you can also use a mask to replace form the original data frame after labelling
df = pd.DataFrame({'A': ['x', np.NaN, 'z'], 'B': [1, 6, 9], 'C': [2, 1, np.NaN]})
A B C
0 x 1 2.0
1 NaN 6 1.0
2 z 9 NaN
original = df
mask = df_1.isnull()
A B C
0 False False False
1 True False False
2 False False True
df = df.astype(str).apply(LabelEncoder().fit_transform)
df.where(~mask, original)
A B C
0 1.0 0 1.0
1 NaN 1 0.0
2 2.0 2 NaN
Hello a little computational hack I did for my own work:
from sklearn.preprocessing import LabelEncoder
import pandas as pd
import numpy as np
a = pd.DataFrame(['A','B','C',np.nan,'D','A'])
le = LabelEncoder()
### fit with the desired col, col in position 0 for this example
fit_by = pd.Series([i for i in a.iloc[:,0].unique() if type(i) == str])
le.fit(fit_by)
### Set transformed col leaving np.NaN as they are
a["transformed"] = fit_by.apply(lambda x: le.transform([x])[0] if type(x) == str else x)
This is my solution, because I was not pleased with the solutions posted here. I needed a LabelEncoder that keeps my missing values as NaN to use an Imputer afterwards. So I have written my own LabelEncoder class. It works with DataFrames.
from sklearn.base import BaseEstimator
from sklearn.base import TransformerMixin
from sklearn.preprocessing import LabelEncoder
class LabelEncoderByCol(BaseEstimator, TransformerMixin):
def __init__(self,col):
#List of column names in the DataFrame that should be encoded
self.col = col
#Dictionary storing a LabelEncoder for each column
self.le_dic = {}
for el in self.col:
self.le_dic[el] = LabelEncoder()
def fit(self,x,y=None):
#Fill missing values with the string 'NaN'
x[self.col] = x[self.col].fillna('NaN')
for el in self.col:
#Only use the values that are not 'NaN' to fit the Encoder
a = x[el][x[el]!='NaN']
self.le_dic[el].fit(a)
return self
def transform(self,x,y=None):
#Fill missing values with the string 'NaN'
x[self.col] = x[self.col].fillna('NaN')
for el in self.col:
#Only use the values that are not 'NaN' to fit the Encoder
a = x[el][x[el]!='NaN']
#Store an ndarray of the current column
b = x[el].to_numpy()
#Replace the elements in the ndarray that are not 'NaN'
#using the transformer
b[b!='NaN'] = self.le_dic[el].transform(a)
#Overwrite the column in the DataFrame
x[el]=b
#return the transformed DataFrame
return x
You can enter a DataFrame, not only a 1-dim Series. with col you can chose the columns that should be encoded.
I would like to here some feedback.
This Module also available in my Github well organized with example.
Please upvoted if you like my solution.
Tks, Idan
class label_encoder_contain_missing_values :
def __init__ (self) :
pass
def categorical_to_numeric (self,dataset):
import numpy as np
import pandas as pd
self.dataset = dataset
self.summary = None
self.table_encoder= {}
for index in self.dataset.columns :
if self.dataset[index].dtypes == 'object' :
column_data_frame = pd.Series(self.dataset[index],name='column').to_frame()
unique_values = pd.Series(self.dataset[index].unique())
i = 0
label_encoder = pd.DataFrame({'value_name':[],'Encode':[]})
while i <= len(unique_values)-1:
if unique_values.isnull()[i] == True :
label_encoder = label_encoder.append({'value_name': unique_values[i],'Encode':np.nan}, ignore_index=True) #np.nan = -1
else:
label_encoder = label_encoder.append({'value_name': unique_values[i],'Encode':i}, ignore_index=True)
i+=1
output = pd.merge(left=column_data_frame,right = label_encoder, how='left',left_on='column',right_on='value_name')
self.summary = output[['column','Encode']].drop_duplicates().reset_index(drop=True)
self.dataset[index] = output.Encode
self.table_encoder.update({index:self.summary})
else :
pass
# ---- Show Encode Table ----- #
print('''\nLabel Encoding completed in Successfully.\n
Next steps: \n
1. To view table_encoder, Execute the follow: \n
for index in table_encoder :
print(f'\\n{index} \\n',table_encoder[index])
2. For inverse, execute the follow : \n
df = label_encoder_contain_missing_values().
inverse_numeric_to_categorical(table_encoder, df) ''')
return self.table_encoder ,self.dataset
def inverse_numeric_to_categorical (self,table_encoder, df):
dataset = df.copy()
for column in table_encoder.keys():
df_column = df[column].to_frame()
output = pd.merge(left=df_column,right = table_encoder[column], how='left',left_on= column,right_on='Encode')#.rename(columns={'column_x' :'encode','column_y':'category'})
df[column]= output.column
print('\nInverse Label Encoding, from categorical to numerical completed in Successfully.\n')
return df
**execute command from categorical to numerical** <br>
table_encoder, df = label_encoder_contain_missing_values().categorical_to_numeric(df)
**execute command from numerical to categorical** <br>
df = label_encoder_contain_missing_values().inverse_numeric_to_categorical(table_encoder, df)
An easy way is this
It is an example of Titanic
LABEL_COL = ["Sex", "Embarked"]
def label(df):
_df = df.copy()
le = LabelEncoder()
for col in LABEL_COL:
# Not NaN index
idx = ~_df[col].isna()
_df.loc[idx, col] \
= le.fit(_df.loc[idx, col]).transform(_df.loc[idx, col])
return _df
The most voted answer by @Kerem has typos, therefore I am posting the corrected and improved answer here:
from sklearn.preprocessing import LabelEncoder
import pandas as pd
import numpy as np
a = pd.DataFrame(['A','B','C',np.nan,'D','A'])
for j in a.columns.values:
le = LabelEncoder()
### fit with the desired col, col in position 0 for this ###example
fit_by = pd.Series([i for i in a[j].unique() if type(i) == str])
le.fit(fit_by)
### Set transformed col leaving np.NaN as they are
a["transformed"] = a[j].apply(lambda x: le.transform([x])[0] if type(x) == str else x)
You can handle missing values by replacing it with string 'NaN'. The category can be obtained by le.transfrom().
le.fit_transform(a.fillna('NaN'))
category = le.transform(['NaN'])
Another solution is for label encoder to ignore missing values.
a = le.fit_transform(a.astype(str))
You can fill the na's by some value and later change the dataframe column type to string to make things work.
from sklearn.preprocessing import LabelEncoder
import pandas as pd
import numpy as np
a = pd.DataFrame(['A','B','C',np.nan,'D','A'])
a.fillna(99)
le = LabelEncoder()
le.fit_transform(a.astype(str))
Following encoder addresses None values in each category.
class MultiColumnLabelEncoder:
def __init__(self):
self.columns = None
self.led = defaultdict(preprocessing.LabelEncoder)
def fit(self, X):
self.columns = X.columns
for col in self.columns:
cat = X[col].unique()
cat = [x if x is not None else "None" for x in cat]
self.led[col].fit(cat)
return self
def fit_transform(self, X):
if self.columns is None:
self.fit(X)
return self.transform(X)
def transform(self, X):
return X.apply(lambda x: self.led[x.name].transform(x.apply(lambda e: e if e is not None else "None")))
def inverse_transform(self, X):
return X.apply(lambda x: self.led[x.name].inverse_transform(x))
Uses Example
df = pd.DataFrame({
'pets': ['cat', 'dog', 'cat', 'monkey', 'dog', 'dog'],
'owner': ['Champ', 'Ron', 'Brick', None, 'Veronica', 'Ron'],
'location': ['San_Diego', 'New_York', 'New_York', 'San_Diego', 'San_Diego',
None]
})
print(df)
location owner pets
0 San_Diego Champ cat
1 New_York Ron dog
2 New_York Brick cat
3 San_Diego None monkey
4 San_Diego Veronica dog
5 None Ron dog
le = MultiColumnLabelEncoder()
le.fit(df)
transformed = le.transform(df)
print(transformed)
location owner pets
0 2 1 0
1 0 3 1
2 0 0 0
3 2 2 2
4 2 4 1
5 1 3 1
inverted = le.inverse_transform(transformed)
print(inverted)
location owner pets
0 San_Diego Champ cat
1 New_York Ron dog
2 New_York Brick cat
3 San_Diego None monkey
4 San_Diego Veronica dog
5 None Ron dog
This function takes a column from a dataframe and return the column where only non-NaNs are label encoded, the rest remains untouched
import pandas as pd
from sklearn.preprocessing import LabelEncoder
def label_encode_column(col):
nans = col.isnull()
nan_lst = []
nan_idx_lst = []
label_lst = []
label_idx_lst = []
for idx, nan in enumerate(nans):
if nan:
nan_lst.append(col[idx])
nan_idx_lst.append(idx)
else:
label_lst.append(col[idx])
label_idx_lst.append(idx)
nan_df = pd.DataFrame(nan_lst, index=nan_idx_lst)
label_df = pd.DataFrame(label_lst, index=label_idx_lst)
label_encoder = LabelEncoder()
label_df = label_encoder.fit_transform(label_df.astype(str))
label_df = pd.DataFrame(label_df, index=label_idx_lst)
final_col = pd.concat([label_df, nan_df])
return final_col.sort_index()
This is how I did it:
import pandas as pd
from sklearn.preprocessing import LabelEncoder
UNKNOWN_TOKEN = '<unknown>'
a = pd.Series(['A','B','C', 'D','A'], dtype=str).unique().tolist()
a.append(UNKNOWN_TOKEN)
le = LabelEncoder()
le.fit_transform(a)
embedding_map = dict(zip(le.classes_, le.transform(le.classes_)))
and when applying to new test data:
test_df = test_df.apply(lambda x: x if x in embedding_map else UNKNOWN_TOKEN)
le.transform(test_df)
I also wanted to contribute my workaround, as I found the others a bit more tedious when working with categorical data which contains missing values
# Create a random dataframe
foo = pd.DataFrame(np.random.randint(0,100,size=(100, 4)), columns=list('ABCD'))
# Randomly intersperse column 'A' with missing data (NaN)
foo['A'][np.random.randint(0,len(foo), size=20)] = np.nan
# Convert this series to string, to simulate our problem
series = foo['A'].astype(str)
# np.nan are converted to the string "nan", mask these out
mask = (series == "nan")
# Apply the LabelEncoder to the unmasked series, replace the masked series with np.nan
series[~mask] = LabelEncoder().fit_transform(series[~mask])
series[mask] = np.nan
foo['A'] = series
This is my attempt!
import numpy as np
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
#Now lets encode the incomplete Cabin feature
titanic_train_le['Cabin'] = le.fit_transform(titanic_train_le['Cabin'].astype(str))
#get nan code for the cabin categorical feature
cabin_nan_code=le.transform(['nan'])[0]
#Now, retrieve the nan values in the encoded data
titanic_train_le['Cabin'].replace(cabin_nan_code,np.nan,inplace=True)
I just created my own encoder which can encode a dataframe at once. Using this class, None is encoded to 0. It can be handy when trying to make sparse matrix. Note that the input dataframe must include categorical columns only.
class DF_encoder():
def __init__(self):
self.mapping = {None : 0}
self.inverse_mapping = {0 : None}
self.all_keys =[]
def fit(self,df:pd.DataFrame):
for col in df.columns:
keys = list(df[col].unique())
self.all_keys += keys
self.all_keys = list(set(self.all_keys))
for i , item in enumerate(start=1 ,iterable=self.all_keys):
if item not in self.mapping.keys():
self.mapping[item] = i
self.inverse_mapping[i] = item
def transform(self,df):
temp_df = pd.DataFrame()
for col in df.columns:
temp_df[col] = df[col].map(self.mapping)
return temp_df
def inverse_transform(self,df):
temp_df = pd.DataFrame()
for col in df.columns:
temp_df[col] = df[col].map(self.inverse_mapping)
return temp_df
I faced the same problem but none of the above worked for me. So I added a new row to the training data consisting only "nan"
