I have a machine learning classification problem with 80% categorical variables. Must I use one hot encoding if I want to use some classifier for the classification? Can i pass the data to a classifier without the encoding?
I am trying to do the following for feature selection:
I read the train file:
num_rows_to_read = 10000 train_small = pd.read_csv("../../dataset/train.csv", nrows=num_rows_to_read)I change the type of the categorical features to 'category':
non_categorial_features = ['orig_destination_distance', 'srch_adults_cnt', 'srch_children_cnt', 'srch_rm_cnt', 'cnt'] for categorical_feature in list(train_small.columns): if categorical_feature not in non_categorial_features: train_small[categorical_feature] = train_small[categorical_feature].astype('category')I use one hot encoding:
train_small_with_dummies = pd.get_dummies(train_small, sparse=True)
The problem is that the 3'rd part often get stuck, although I am using a strong machine.
Thus, without the one hot encoding I can't do any feature selection, for determining the importance of the features.
What do you recommend?
Approach 1: You can use get_dummies on pandas dataframe.
Example 1:
import pandas as pd
s = pd.Series(list('abca'))
pd.get_dummies(s)
Out[]:
a b c
0 1.0 0.0 0.0
1 0.0 1.0 0.0
2 0.0 0.0 1.0
3 1.0 0.0 0.0
Example 2:
The following will transform a given column into one hot. Use prefix to have multiple dummies.
import pandas as pd
df = pd.DataFrame({
'A':['a','b','a'],
'B':['b','a','c']
})
df
Out[]:
A B
0 a b
1 b a
2 a c
# Get one hot encoding of columns B
one_hot = pd.get_dummies(df['B'])
# Drop column B as it is now encoded
df = df.drop('B',axis = 1)
# Join the encoded df
df = df.join(one_hot)
df
Out[]:
A a b c
0 a 0 1 0
1 b 1 0 0
2 a 0 0 1
Approach 2: Use Scikit-learn
Given a dataset with three features and four samples, we let the encoder find the maximum value per feature and transform the data to a binary one-hot encoding.
>>> from sklearn.preprocessing import OneHotEncoder
>>> enc = OneHotEncoder()
>>> enc.fit([[0, 0, 3], [1, 1, 0], [0, 2, 1], [1, 0, 2]])
OneHotEncoder(categorical_features='all', dtype=<class 'numpy.float64'>,
handle_unknown='error', n_values='auto', sparse=True)
>>> enc.n_values_
array([2, 3, 4])
>>> enc.feature_indices_
array([0, 2, 5, 9], dtype=int32)
>>> enc.transform([[0, 1, 1]]).toarray()
array([[ 1., 0., 0., 1., 0., 0., 1., 0., 0.]])
Here is the link for this example: http://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OneHotEncoder.html
You can do it with numpy.eye and a using the array element selection mechanism:
import numpy as np
nb_classes = 6
data = [[2, 3, 4, 0]]
def indices_to_one_hot(data, nb_classes):
"""Convert an iterable of indices to one-hot encoded labels."""
targets = np.array(data).reshape(-1)
return np.eye(nb_classes)[targets]
The the return value of indices_to_one_hot(nb_classes, data) is now
array([[[ 0., 0., 1., 0., 0., 0.],
[ 0., 0., 0., 1., 0., 0.],
[ 0., 0., 0., 0., 1., 0.],
[ 1., 0., 0., 0., 0., 0.]]])
The .reshape(-1) is there to make sure you have the right labels format (you might also have [[2], [3], [4], [0]]).
Firstly, easiest way to one hot encode: use Sklearn.
http://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OneHotEncoder.html
Secondly, I don't think using pandas to one hot encode is that simple (unconfirmed though)
Creating dummy variables in pandas for python
Lastly, is it necessary for you to one hot encode? One hot encoding exponentially increases the number of features, drastically increasing the run time of any classifier or anything else you are going to run. Especially when each categorical feature has many levels. Instead you can do dummy coding.
Using dummy encoding usually works well, for much less run time and complexity. A wise prof once told me, 'Less is More'.
Here's the code for my custom encoding function if you want.
from sklearn.preprocessing import LabelEncoder
#Auto encodes any dataframe column of type category or object.
def dummyEncode(df):
columnsToEncode = list(df.select_dtypes(include=['category','object']))
le = LabelEncoder()
for feature in columnsToEncode:
try:
df[feature] = le.fit_transform(df[feature])
except:
print('Error encoding '+feature)
return df
EDIT: Comparison to be clearer:
One-hot encoding: convert n levels to n-1 columns.
Index Animal Index cat mouse
1 dog 1 0 0
2 cat --> 2 1 0
3 mouse 3 0 1
You can see how this will explode your memory if you have many different types (or levels) in your categorical feature. Keep in mind, this is just ONE column.
Dummy Coding:
Index Animal Index Animal
1 dog 1 0
2 cat --> 2 1
3 mouse 3 2
Convert to numerical representations instead. Greatly saves feature space, at the cost of a bit of accuracy.
One hot encoding with pandas is very easy:
def one_hot(df, cols):
"""
@param df pandas DataFrame
@param cols a list of columns to encode
@return a DataFrame with one-hot encoding
"""
for each in cols:
dummies = pd.get_dummies(df[each], prefix=each, drop_first=False)
df = pd.concat([df, dummies], axis=1)
return df
EDIT:
Another way to one_hot using sklearn's LabelBinarizer :
from sklearn.preprocessing import LabelBinarizer
label_binarizer = LabelBinarizer()
label_binarizer.fit(all_your_labels_list) # need to be global or remembered to use it later
def one_hot_encode(x):
"""
One hot encode a list of sample labels. Return a one-hot encoded vector for each label.
: x: List of sample Labels
: return: Numpy array of one-hot encoded labels
"""
return label_binarizer.transform(x)
You can use numpy.eye function.
import numpy as np
def one_hot_encode(x, n_classes):
"""
One hot encode a list of sample labels. Return a one-hot encoded vector for each label.
: x: List of sample Labels
: return: Numpy array of one-hot encoded labels
"""
return np.eye(n_classes)[x]
def main():
list = [0,1,2,3,4,3,2,1,0]
n_classes = 5
one_hot_list = one_hot_encode(list, n_classes)
print(one_hot_list)
if __name__ == "__main__":
main()
Result
D:\Desktop>python test.py
[[ 1. 0. 0. 0. 0.]
[ 0. 1. 0. 0. 0.]
[ 0. 0. 1. 0. 0.]
[ 0. 0. 0. 1. 0.]
[ 0. 0. 0. 0. 1.]
[ 0. 0. 0. 1. 0.]
[ 0. 0. 1. 0. 0.]
[ 0. 1. 0. 0. 0.]
[ 1. 0. 0. 0. 0.]]
Much easier to use Pandas for basic one-hot encoding. If you're looking for more options you can use scikit-learn.
For basic one-hot encoding with Pandas you simply pass your data frame into the get_dummies function.
For example, if I have a dataframe called imdb_movies:
...and I want to one-hot encode the Rated column, I simply do this:
pd.get_dummies(imdb_movies.Rated)
This returns a new dataframe with a column for every "level" of rating that exists, along with either a 1 or 0 specifying the presence of that rating for a given observation.
Usually, we want this to be part of the original dataframe. In this case we simply attach our new dummy coded frame onto the original frame using "column-binding.
We can column-bind by using Pandas concat function:
rated_dummies = pd.get_dummies(imdb_movies.Rated)
pd.concat([imdb_movies, rated_dummies], axis=1)
We can now run analysis on our full dataframe.
SIMPLE UTILITY FUNCTION
I would recommend making yourself a utility function to do this quickly:
def encode_and_bind(original_dataframe, feature_to_encode):
dummies = pd.get_dummies(original_dataframe[[feature_to_encode]])
res = pd.concat([original_dataframe, dummies], axis=1)
return(res)
Usage:
encode_and_bind(imdb_movies, 'Rated')
Result:
Also, as per @pmalbu comment, if you would like the function to remove the original feature_to_encode then use this version:
def encode_and_bind(original_dataframe, feature_to_encode):
dummies = pd.get_dummies(original_dataframe[[feature_to_encode]])
res = pd.concat([original_dataframe, dummies], axis=1)
res = res.drop([feature_to_encode], axis=1)
return(res)
One-hot encoding requires bit more than converting the values to indicator variables. Typically ML process requires you to apply this coding several times to validation or test data sets and applying the model you construct to real-time observed data. You should store the mapping (transform) that was used to construct the model. A good solution would use the DictVectorizer or LabelEncoder (followed by get_dummies. Here is a function that you can use:
def oneHotEncode2(df, le_dict = {}):
if not le_dict:
columnsToEncode = list(df.select_dtypes(include=['category','object']))
train = True;
else:
columnsToEncode = le_dict.keys()
train = False;
for feature in columnsToEncode:
if train:
le_dict[feature] = LabelEncoder()
try:
if train:
df[feature] = le_dict[feature].fit_transform(df[feature])
else:
df[feature] = le_dict[feature].transform(df[feature])
df = pd.concat([df,
pd.get_dummies(df[feature]).rename(columns=lambda x: feature + '_' + str(x))], axis=1)
df = df.drop(feature, axis=1)
except:
print('Error encoding '+feature)
#df[feature] = df[feature].convert_objects(convert_numeric='force')
df[feature] = df[feature].apply(pd.to_numeric, errors='coerce')
return (df, le_dict)
This works on a pandas dataframe and for each column of the dataframe it creates and returns a mapping back. So you would call it like this:
train_data, le_dict = oneHotEncode2(train_data)
Then on the test data, the call is made by passing the dictionary returned back from training:
test_data, _ = oneHotEncode2(test_data, le_dict)
An equivalent method is to use DictVectorizer. A related post on the same is on my blog. I mention it here since it provides some reasoning behind this approach over simply using get_dummies post (disclosure: this is my own blog).
Here is a solution using DictVectorizer and the Pandas DataFrame.to_dict('records') method.
>>> import pandas as pd
>>> X = pd.DataFrame({'income': [100000,110000,90000,30000,14000,50000],
'country':['US', 'CAN', 'US', 'CAN', 'MEX', 'US'],
'race':['White', 'Black', 'Latino', 'White', 'White', 'Black']
})
>>> from sklearn.feature_extraction import DictVectorizer
>>> v = DictVectorizer()
>>> qualitative_features = ['country','race']
>>> X_qual = v.fit_transform(X[qualitative_features].to_dict('records'))
>>> v.vocabulary_
{'country=CAN': 0,
'country=MEX': 1,
'country=US': 2,
'race=Black': 3,
'race=Latino': 4,
'race=White': 5}
>>> X_qual.toarray()
array([[ 0., 0., 1., 0., 0., 1.],
[ 1., 0., 0., 1., 0., 0.],
[ 0., 0., 1., 0., 1., 0.],
[ 1., 0., 0., 0., 0., 1.],
[ 0., 1., 0., 0., 0., 1.],
[ 0., 0., 1., 1., 0., 0.]])
pandas as has inbuilt function "get_dummies" to get one hot encoding of that particular column/s.
one line code for one-hot-encoding:
df=pd.concat([df,pd.get_dummies(df['column name'],prefix='column name')],axis=1).drop(['column name'],axis=1)
I know I'm late to this party, but the simplest way to hot encode a dataframe in an automated way is to use this function:
def hot_encode(df):
obj_df = df.select_dtypes(include=['object'])
return pd.get_dummies(df, columns=obj_df.columns).values
You can pass the data to catboost classifier without encoding. Catboost handles categorical variables itself by performing one-hot and target expanding mean encoding.
To add to other questions, let me provide how I did it with a Python 2.0 function using Numpy:
def one_hot(y_):
# Function to encode output labels from number indexes
# e.g.: [[5], [0], [3]] --> [[0, 0, 0, 0, 0, 1], [1, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0]]
y_ = y_.reshape(len(y_))
n_values = np.max(y_) + 1
return np.eye(n_values)[np.array(y_, dtype=np.int32)] # Returns FLOATS
The line n_values = np.max(y_) + 1 could be hard-coded for you to use the good number of neurons in case you use mini-batches for example.
Demo project/tutorial where this function has been used: https://github.com/guillaume-chevalier/LSTM-Human-Activity-Recognition
This works for me:
pandas.factorize( ['B', 'C', 'D', 'B'] )[0]
Output:
[0, 1, 2, 0]
I used this in my acoustic model: probably this helps in ur model.
def one_hot_encoding(x, n_out):
x = x.astype(int)
shape = x.shape
x = x.flatten()
N = len(x)
x_categ = np.zeros((N,n_out))
x_categ[np.arange(N), x] = 1
return x_categ.reshape((shape)+(n_out,))
It can and it should be easy as :
class OneHotEncoder:
def __init__(self,optionKeys):
length=len(optionKeys)
self.__dict__={optionKeys[j]:[0 if i!=j else 1 for i in range(length)] for j in range(length)}
Usage :
ohe=OneHotEncoder(["A","B","C","D"])
print(ohe.A)
print(ohe.D)
Here i tried with this approach :
import numpy as np
#converting to one_hot
def one_hot_encoder(value, datal):
datal[value] = 1
return datal
def _one_hot_values(labels_data):
encoded = [0] * len(labels_data)
for j, i in enumerate(labels_data):
max_value = [0] * (np.max(labels_data) + 1)
encoded[j] = one_hot_encoder(i, max_value)
return np.array(encoded)
来源：https://stackoverflow.com/questions/37292872/how-can-i-one-hot-encode-in-python