一个在numpy中使用矢量化并在pandas中应用的简单示例:

import numpy as np

a = np.array(['male','female','female','male'])

#define function
onehot_function = lambda x: 1.0 if (x=='male') else 0.0

onehot_a = np.vectorize(onehot_function)(a)

print(onehot_a)
# [1., 0., 0., 1.]

# -----------------------------------------

import pandas as pd

s = pd.Series(['male','female','female','male'])
onehot_s = s.apply(onehot_function)

print(onehot_s)
# 0    1.0
# 1    0.0
# 2    0.0
# 3    1.0
# dtype: float64

您可以将数据传递给catboost分类器而不进行编码。Catboost通过执行单热和目标扩展平均编码来处理分类变量本身。

Pandas as有内置函数“get_dummies”来获取特定列/秒的热编码。

单热编码的一行代码:

df=pd.concat([df,pd.get_dummies(df['column name'],prefix='column name')],axis=1).drop(['column name'],axis=1)

我在我的声学模型中使用了这个: 也许这对你的模型有帮助。

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,))

简短的回答

这里有一个函数，可以在不使用numpy、pandas或其他包的情况下进行一次性编码。它接受一个整数、布尔值或字符串(也可能是其他类型)的列表。

import typing


def one_hot_encode(items: list) -> typing.List[list]:
    results = []
    # find the unique items (we want to unique items b/c duplicate items will have the same encoding)
    unique_items = list(set(items))
    # sort the unique items
    sorted_items = sorted(unique_items)
    # find how long the list of each item should be
    max_index = len(unique_items)

    for item in items:
        # create a list of zeros the appropriate length
        one_hot_encoded_result = [0 for i in range(0, max_index)]
        # find the index of the item
        one_hot_index = sorted_items.index(item)
        # change the zero at the index from the previous line to a one
        one_hot_encoded_result[one_hot_index] = 1
        # add the result
        results.append(one_hot_encoded_result)

    return results

例子:

one_hot_encode([2, 1, 1, 2, 5, 3])

# [[0, 1, 0, 0],
#  [1, 0, 0, 0],
#  [1, 0, 0, 0],
#  [0, 1, 0, 0],
#  [0, 0, 0, 1],
#  [0, 0, 1, 0]]

one_hot_encode([True, False, True])

# [[0, 1], [1, 0], [0, 1]]

one_hot_encode(['a', 'b', 'c', 'a', 'e'])

# [[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], [1, 0, 0, 0], [0, 0, 0, 1]]

长(er)回答

I know there are already a lot of answers to this question, but I noticed two things. First, most of the answers use packages like numpy and/or pandas. And this is a good thing. If you are writing production code, you should probably be using robust, fast algorithms like those provided in the numpy/pandas packages. But, for the sake of education, I think someone should provide an answer which has a transparent algorithm and not just an implementation of someone else's algorithm. Second, I noticed that many of the answers do not provide a robust implementation of one-hot encoding because they do not meet one of the requirements below. Below are some of the requirements (as I see them) for a useful, accurate, and robust one-hot encoding function:

单热编码函数必须:

处理各种类型的列表(例如，整数，字符串，浮点数等)作为输入 处理带有重复项的输入列表 返回与输入相对应(与输入顺序相同)的列表的列表 返回一个列表的列表，其中每个列表都尽可能短

我测试了这个问题的许多答案，其中大多数都未能满足上述要求之一。

单热编码需要比将值转换为指示器变量更多的内容。通常ML过程要求您多次将此编码应用于验证或测试数据集，并将您构建的模型应用于实时观察数据。您应该存储用于构造模型的映射(转换)。一个好的解决方案是使用DictVectorizer或LabelEncoder(后面跟着get_dummies)。下面是一个你可以使用的函数:

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)

这适用于pandas数据框架，它为数据框架的每一列创建并返回一个映射。所以你可以这样称呼它:

train_data, le_dict = oneHotEncode2(train_data)

然后在测试数据上，通过传递训练返回的字典进行调用:

test_data, _ = oneHotEncode2(test_data, le_dict)

一个等效的方法是使用DictVectorizer。我的博客上有一篇相关的文章。我在这里提到它是因为它为这种方法提供了一些理由，而不是简单地使用get_dummies post(披露:这是我自己的博客)。