http://blog.csdn.net/pipisorry/article/details/52247679
本blog内容有特征预处理(标准化、归一化、正则化、特征二值化、缺失值处理)和标签label预处理(label二值化、multi-label多值化)。
[数据标准化/归一化normalization ]
[均值、方差与协方差矩阵 ]
[矩阵论:向量范数和矩阵范数 ]
Note: 一定要注意归一化是归一化什么,归一化features还是samples。
Standardization: mean removal and variance scaling
数据标准化:当单个特征的样本取值相差甚大或明显不遵从高斯正态分布时,标准化表现的效果较差。实际操作中,经常忽略特征数据的分布形状,移除每个特征均值,划分离散特征的标准差,从而等级化,进而实现数据中心化。
Note: test set要和training set做相同的预处理操作(standardization、data transformation、etc)。
[数据标准化/归一化normalization ]
from sklearn import preprocessing
preprocessing.scale(X)
def scale(X, axis=0, with_mean=True, with_std=True, copy=True)
注意,scikit-learn中assume that all features are centered around zero and have variance in the same order.同时这个默认操作是对features进行的(如mean removal),所以操作都是针对axis=0的操作,如果数据不是这样的要注意!公式为:(X-X_mean)/X_std 计算时对每个属性/每列分别进行。
参数解释:
X:{array-like, sparse matrix} 数组或者矩阵,一维的数据都可以(但是在0.19版本后一维的数据会报错了!)
axis:int类型,初始值为0,axis用来计算均值 means 和标准方差 standard deviations. 如果是0,则单独的标准化每个特征(列),如果是1,则标准化每个观测样本(行)。
with_mean: boolean类型,默认为True,表示将数据均值规范到0
with_std: boolean类型,默认为True,表示将数据方差规范到1
这种标准化相当于z-score 标准化(zero-mean normalization)
[sklearn.preprocessing
.scale]
scale标准化示例
>>> from sklearn import preprocessing
>>> import numpy as np
>>> X = np.array([[ 1., -1., 2.],
... [ 2., 0., 0.],
... [ 0., 1., -1.]])
>>> X_scaled = preprocessing.scale(X)
>>> X_scaled
array([[ 0. ..., -1.22..., 1.33...],
[ 1.22..., 0. ..., -0.26...],
[-1.22..., 1.22..., -1.06...]])
对于一维数据的一种可能的处理:先转换成二维,再在结果中转换为一维
cn = preprocessing.scale([[p] for _, _, p in cn]).reshape(-1)
转换后的数据有0均值(zero mean)和单位方差(unit variance,方差为1)
>>> X_scaled.mean(axis=0)
array([ 0., 0., 0.])
>>> X_scaled.std(axis=0)
array([ 1., 1., 1.])
一般我们的标准化先在训练集上进行,在测试集上也应该做同样mean和variance的标准化,这样就应该将训练集上的标准化参数保存下来。
The preprocessing module further provides a utility class StandardScaler that implements the Transformer API to computethe mean and standard deviation on a training set so as to beable to later reapply the same transformation on the testing set.This class is hence suitable for use in the early steps of a sklearn.pipeline.Pipeline:
>>> scaler = preprocessing.StandardScaler().fit(X)
>>> scaler
StandardScaler(copy=True, with_mean=True, with_std=True)
>>> scaler.mean_
array([ 1. ..., 0. ..., 0.33...])
>>> scaler.scale_
array([ 0.81..., 0.81..., 1.24...])
>>> scaler.transform(X)
array([[ 0. ..., -1.22..., 1.33...],
[ 1.22..., 0. ..., -0.26...],
[-1.22..., 1.22..., -1.06...]])
The scaler instance can then be used on new data to transform it thesame way it did on the training set:
>>> scaler.transform([[-1., 1., 0.]])
array([[-2.44..., 1.22..., -0.26...]])
It is possible to disable either centering or scaling by eitherpassing with_mean=False or with_std=False to the constructorof StandardScaler.[StandardScaler]
[Standardization, or mean removal and variance scaling]
StandardScaler示例
def preprocess():
if not os.path.exists(os.path.join(DIR, train_file1)) or not os.path.exists(os.path.join(DIR, test_file1)) or 0:
xy = np.loadtxt(os.path.join(DIR, train_file), delimiter=',', dtype=float)
x, y = xy[:, 0:-1], xy[:, -1]
scaler = preprocessing.StandardScaler().fit(x)
xy = np.hstack([scaler.transform(x), y])
np.savetxt(os.path.join(DIR, train_file1), xy, fmt='%.7f')
x_test = np.loadtxt(os.path.join(DIR, test_file), delimiter=',', dtype=float)
x_test = scaler.transform(x_test)
np.savetxt(os.path.join(DIR, test_file1), x_test, fmt='%.7f')
else:
print('data loading...')
xy = np.loadtxt(os.path.join(DIR, train_file1), dtype=float)
x_test = np.loadtxt(os.path.join(DIR, test_file1), dtype=float)
return xy[:, 0:-1], xy[:, -1], x_test
Note:
pipeline能简化该过程( See Pipeline and FeatureUnion: combining estimators ,翻译后的文章:http://www.voidcn.com/blog/mmc2015/article/p-3379231.html):
>>> from sklearn.pipeline import make_pipeline
>>> clf = make_pipeline(preprocessing.StandardScaler(), svm.SVC(C=1))
>>> cross_validation.cross_val_score(clf, iris.data, iris.target, cv=cv)
...
array([ 0.97..., 0.93..., 0.95...])
将属性缩放到一个指定的最大值和最小值(通常是1-0)之间,这可以通过preprocessing.MinMaxScaler类来实现。
使用这种方法的目的包括:
1、对于方差非常小的属性可以增强其稳定性;
2、维持稀疏矩阵中为0的条目。
min_max_scaler = preprocessing.MinMaxScaler()
X_minMax = min_max_scaler.fit_transform(X)
sklearn.preprocessing.robust_scale(X, axis=0, with_centering=True, with_scaling=True, quantile_range=(25.0, 75.0), copy=True)
Center to the median and component wise scaleaccording to the interquartile range.
[Scaling data with outliers]
其它
[Scaling sparse data
Centering kernel matrices]
Constructs a transformer from an arbitrary callable.
lz自定义了一个归一化函数:大于某个THRESHOLD时其属于1的概率值要大于0.5,小于THRESHOLD时概率值小于0.5,接近最大值时其概率值越接近1,接近最小值时其概率值越接近0。相当于min-max归一化的一点改进吧。
from sklearn.preprocessing import FunctionTransformer
import numpy as np
def scalerFunc(x, maxv, minv, THRESHOLD=200):
'''
:param x: (n_samples, n_features)!!
'''
label = x >= THRESHOLD
result = 0.5 * (1 + (x - THRESHOLD) * (label / (maxv - THRESHOLD) + (label - 1) / (minv - THRESHOLD)))
# print(result)
return result
x = np.array([100, 150, 201, 250, 300]).reshape(-1, 1)
scaler = FunctionTransformer(func=scalerFunc, kw_args={'maxv': x.max(), 'minv': x.min()}).fit(x)
print(scaler.transform(x))
[[ 0. ] [ 0.25 ] [ 0.505] [ 0.75 ] [ 1. ]]
Note: 自定义函数的参数由FunctionTransformer中的kw_args指定,是字典类型,key必须是字符串。
[preprocessing.FunctionTransformer([func, ...])]
[sklearn.preprocessing: Preprocessing and Normalization¶]
正则化的过程是将每个样本缩放到单位范数(每个样本的范数为1),如果要使用如二次型(点积)或者其它核方法计算两个样本之间的相似性这个方法会很有用。
该方法是文本分类和聚类分析中经常使用的向量空间模型(Vector Space Model)的基础.
Normalization主要思想是对每个样本计算其p-范数,然后对该样本中每个元素除以该范数,这样处理的结果是使得每个处理后样本的p-范数(l1-norm,l2-norm)等于1。
Normalization is the process of scaling individual samples to haveunit norm.This process can be useful if you plan to use a quadratic formsuch as the dot-product or any other kernel to quantify the similarityof any pair of samples.This assumption is the base of the Vector Space Model often used in textclassification and clustering contexts.
def normalize(X, norm='l2', axis=1, copy=True)
注意,这个操作是对所有样本(而不是features)进行的,也就是将每个样本的值除以这个样本的Li范数。所以这个操作是针对axis=1进行的。
>>> X = [[ 1., -1., 2.],
... [ 2., 0., 0.],
... [ 0., 1., -1.]]
>>> X_normalized = preprocessing.normalize(X, norm='l2')
>>> X_normalized
array([[ 0.40..., -0.40..., 0.81...],
[ 1. ..., 0. ..., 0. ...],
[ 0. ..., 0.70..., -0.70...]])
[Normalization]
皮皮blog
由于不同的原因,许多现实中的数据集都包含有缺失值,要么是空白的,要么使用NaNs或者其它的符号替代。这些数据无法直接使用scikit-learn分类器直接训练,所以需要进行处理。幸运地是,sklearn中的Imputer类提供了一些基本的方法来处理缺失值,如使用均值、中位值或者缺失值所在列中频繁出现的值来替换。
Imputer类同样支持稀疏矩阵。
>>> import numpy as np
>>> from sklearn.preprocessing import Imputer
>>> imp = Imputer(missing_values='NaN', strategy='mean', axis=0)
>>> imp.fit([[1, 2], [np.nan, 3], [7, 6]])
Imputer(axis=0, copy=True, missing_values='NaN', strategy='mean', verbose=0)
>>> X = [[np.nan, 2], [6, np.nan], [7, 6]]
>>> print(imp.transform(X))
[[ 4. 2. ]
[ 6. 3.666...]
[ 7. 6. ]]
不过lz更倾向于使用pandas进行数据的这种处理[pandas小记:pandas高级功能 ]。
[Imputation of missing values]
[Generating polynomial features]
[Custom transformers]
皮皮blog
Binarize data (set feature values to 0 or 1) according to a threshold. LabelBinarizer
is a utility class to help create a label indicator matrix from a list of multi-class labels. 特征的二值化主要是为了将数据特征转变成boolean变量。
sklearn.preprocessing.Binarizer函数可以设定一个阈值,结果数据值大于阈值的为1,小于阈值的为0。
>>> X = [[ 1., -1., 2.],
... [ 2., 0., 0.],
... [ 0., 1., -1.]]
>>> binarizer = preprocessing.Binarizer().fit(X) # fit does nothing
>>> binarizer
Binarizer(copy=True, threshold=0.0)
>>> binarizer.transform(X)
array([[ 1., 0., 1.],
[ 1., 0., 0.],
[ 0., 1., 0.]])
[Binarization]
Transform between iterable of iterables and a multilabel format
fit字典中所有的字
mlb = MultiLabelBinarizer()
with open(os.path.join(DATADIR, 'vocab.tags.txt'), 'r', encoding='utf-8') as f:
mlb.fit([[l.strip() for l in f.readlines()]])
类数目
mlb.classes_.size
所有类名的ndarray
mlb.classes_
转换class_names为class_ids
transform(self, y)
y : iterable of iterables. 是一个可迭代对象就可以,当然其中的数据需要是mlb.classes_中的数据。返回一个二维的 (n_samples, n_classes) 的multi-hot表示。
转换class_ids为class_names
inverse_transform(self, yt)
其中参数:yt : array or sparse matrix of shape (n_samples, n_classes) .A matrix containing only 1s ands 0s. 必须是一个二维的有shape参数的ndarray或者tensor具体值(所以如果只是一个一维数据需要先转成np.array([ndarray_data])或者tf.expand_dims(tensor_data, 0)),且其中的数据不能是logits,而应该是0,1值的ids。
返回:y : list of tuples. The set of labels for each sample such that y[i] consists of classes_[j] for each yt[i, j] == 1. 返回的是一个一维列表,其中的元素为label的tuple(因为可能是multi-label)。
输出示例
from sklearn.preprocessing import MultiLabelBinarizer
import numpy as np
mlb = MultiLabelBinarizer()
ids = mlb.fit_transform([('a', 'b'), ('大', '小'), ('大',), ('左右', '晨')])
ids = mlb.transform(['a', '小'])
labels1 = mlb.inverse_transform(ids)
labels2 = mlb.inverse_transform(np.array([[0, 0, 0, 1, 0, 0]]))
print(ids)
print(mlb.classes_.size)
print(mlb.classes_)
print(ids)
print(labels1)
print(labels2)
[[1 0 0 0 0 0]
[0 0 0 1 0 0]]
6
['a' 'b' '大' '小' '左右' '晨']
[[1 0 0 0 0 0]
[0 0 0 1 0 0]]
[('a',), ('小',)]
[('小',)]
[preprocessing.MultiLabelBinarizer
([classes, …])]
[Encoding categorical features]
皮皮blog
from: http://blog.csdn.net/pipisorry/article/details/52247679
ref: [sklearn.preprocessing
: Preprocessing and Normalization¶]
[Preprocessing data]