我有一个80%类别变量的机器学习分类问题。如果我想使用一些分类器进行分类,我必须使用一个热编码吗?我可以将数据传递给分类器而不进行编码吗?

我试图做以下的特征选择:

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)

问题是,第三部分经常卡住,尽管我使用的是一个强大的机器。

因此,如果没有一个热编码,我就无法进行任何特征选择,以确定特征的重要性。

你有什么建议吗?


当前回答

试试这个:

!pip install category_encoders
import category_encoders as ce

categorical_columns = [...the list of names of the columns you want to one-hot-encode ...]
encoder = ce.OneHotEncoder(cols=categorical_columns, use_cat_names=True)
df_train_encoded = encoder.fit_transform(df_train_small)

df_encoded.head ()

生成的数据框架df_train_encoded与原始数据框架相同,但是分类特征现在被它们的单热编码版本所取代。

更多关于category_encoders的信息请点击这里。

其他回答

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)。

熊猫的热编码非常简单:

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

编辑:

使用sklearn的LabelBinarizer实现one_hot的另一种方法:

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)

为了补充其他问题,让我提供如何使用Numpy使用Python 2.0函数:

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

行n_values = np.max(y_) + 1可以硬编码,以便在使用小批量的情况下使用足够数量的神经元。

使用此函数的演示项目/教程: https://github.com/guillaume-chevalier/LSTM-Human-Activity-Recognition

扩展@Martin Thoma的答案

def one_hot_encode(y):
    """Convert an iterable of indices to one-hot encoded labels."""
    y = y.flatten() # Sometimes not flattened vector is passed e.g (118,1) in these cases
    # the function ends up creating a tensor e.g. (118, 2, 1). flatten removes this issue
    nb_classes = len(np.unique(y)) # get the number of unique classes
    standardised_labels = dict(zip(np.unique(y), np.arange(nb_classes))) # get the class labels as a dictionary
    # which then is standardised. E.g imagine class labels are (4,7,9) if a vector of y containing 4,7 and 9 is
    # directly passed then np.eye(nb_classes)[4] or 7,9 throws an out of index error.
    # standardised labels fixes this issue by returning a dictionary;
    # standardised_labels = {4:0, 7:1, 9:2}. The values of the dictionary are mapped to keys in y array.
    # standardised_labels also removes the error that is raised if the labels are floats. E.g. 1.0; element
    # cannot be called by an integer index e.g y[1.0] - throws an index error.
    targets = np.vectorize(standardised_labels.get)(y) # map the dictionary values to array.
    return np.eye(nb_classes)[targets]

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。