数据集分割方法

在机器学习建模过程中，通行的做法通常是将数据集分为训练集和测试集。测试集是与训练独立的数据，完全不参与训练，用于最终模型的评估。
在训练过程中，经常会出现过拟合的问题，就是模型可以很好的匹配训练数据，却不能很好在预测训练集外的数据。如果此时就使用测试数据来调整模型参数，就相当于在训练时已知部分测试数据的信息，会影响最终评估结果的准确性。通常的做法是在训练数据再中分出一部分做为验证(Validation)数据，用来评估模型的训练效果。
验证数据取自训练数据，但不参与训练，这样可以相对客观的评估模型对于训练集之外数据的匹配程度。模型在验证数据中的评估常用的是交叉验证，又称循环验证。它将原始数据分成K组(K-Fold)，将每个子集数据分别做一次验证集，其余的K-1组子集数据作为训练集，这样会得到K个模型。这K个模型分别在验证集中评估结果，最后的误差MSE(Mean Squared Error)加和平均就得到交叉验证误差。交叉验证有效利用了有限的数据，并且评估结果能够尽可能接近模型在测试集上的表现，可以做为模型优化的指标使用。

一、随机划分

train_test_split是最简单的数据集分割函数，功能是将样本按比例分割。

# coding = utf-8
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn import datasets

iris = datasets.load_iris()
X = iris.data
y = iris.target

x_train,x_test,y_train,y_test = train_test_split(X, y, test_size=0.2,
                                                 stratify=y,  # 按照标签来分层采样
                                                 shuffle=True, # 是否先打乱数据的顺序再划分
                                                 random_state=1)   # 控制将样本随机打乱
x_train,x_valid,y_train,y_valid = train_test_split(x_train, y_train, test_size=0.4,
                                                 stratify=y_train,
                                                 shuffle=True,
                                                 random_state=1) 

然而这种方式并不是很好，有两大缺点：一是浪费数据，二是容易过拟合且矫正方式不方便

二、K折划分

1、KFold

K-Fold是最简单的K折交叉，n-split就是K值，shuffle指是否对数据洗牌，random_state为随机种子
K值的选取会影响bias和viriance。K越大，每次投入的训练集的数据越多，模型的Bias越小。但是K越大，又意味着每一次选取的训练集之前的相关性越大，而这种大相关性会导致最终的test error具有更大的Variance。一般来说，根据经验我们一般选择k=5或10。

# coding = utf-8
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.model_selection import KFold
from sklearn import datasets
from sklearn.ensemble import GradientBoostingClassifier as GBDT
from sklearn.metrics import precision_score

iris = datasets.load_iris()
X = iris.data
y = iris.target

x_train,x_test,y_train,y_test = train_test_split(X, y, test_size=0.2,
                                                 stratify=y,  # 按照标签来分层采样
                                                 shuffle=True, # 是否先打乱数据的顺序再划分
                                                 random_state=1)   # 控制将样本随机打乱

clf = GBDT(n_estimators=100)
precision_scores = []

kf = KFold(n_splits=5, random_state=0, shuffle=False)
for train_index, valid_index in kf.split(x_train, y_train):
    x_train, x_valid = X[train_index], X[valid_index]
    y_train, y_valid = y[train_index], y[valid_index]
    clf.fit(x_train, y_train)
    y_pred = clf.predict(x_valid)
    precision_scores.append(precision_score(y_valid, y_pred, average='micro'))

print('Precision', np.mean(precision_scores))

2、StratifiedKFold

StratifiedKFold用法类似Kfold，但是他是分层采样，确保训练集、验证集中各类别样本的比例与原始数据集中相同。

# coding = utf-8
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.model_selection import KFold
from sklearn import datasets
from sklearn.ensemble import GradientBoostingClassifier as GBDT
from sklearn.metrics import precision_score

iris = datasets.load_iris()
X = iris.data
y = iris.target

x_train,x_test,y_train,y_test = train_test_split(X, y, test_size=0.2,
                                                 stratify=y,  # 按照标签来分层采样
                                                 shuffle=True, # 是否先打乱数据的顺序再划分
                                                 random_state=1)   # 控制将样本随机打乱

clf = GBDT(n_estimators=100)
precision_scores = []

kf = StratifiedKFold(n_splits=5, random_state=0, shuffle=False)
for train_index, valid_index in kf.split(x_train, y_train):
    x_train, x_valid = X[train_index], X[valid_index]
    y_train, y_valid = y[train_index], y[valid_index]
    clf.fit(x_train, y_train)
    y_pred = clf.predict(x_valid)
    precision_scores.append(precision_score(y_valid, y_pred, average='micro'))

print('Precision', np.mean(precision_scores))

3、StratifiedShuffleSplit

StratifiedShuffleSplit允许设置设置train/valid对中train和valid所占的比例

# coding = utf-8
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.model_selection import StratifiedShuffleSplit
from sklearn import datasets
from sklearn.ensemble import GradientBoostingClassifier as GBDT
from sklearn.metrics import precision_score

iris = datasets.load_iris()
X = iris.data
y = iris.target

x_train,x_test,y_train,y_test = train_test_split(X, y, test_size=0.2,
                                                 stratify=y,  # 按照标签来分层采样
                                                 shuffle=True, # 是否先打乱数据的顺序再划分
                                                 random_state=1)   # 控制将样本随机打乱

clf = GBDT(n_estimators=100)
precision_scores = []

kf = StratifiedShuffleSplit(n_splits=10, train_size=0.6, test_size=0.4, random_state=0)
for train_index, valid_index in kf.split(x_train, y_train):
    x_train, x_valid = X[train_index], X[valid_index]
    y_train, y_valid = y[train_index], y[valid_index]
    clf.fit(x_train, y_train)
    y_pred = clf.predict(x_valid)
    precision_scores.append(precision_score(y_valid, y_pred, average='micro'))

print('Precision', np.mean(precision_scores))

其他的方法如RepeatedStratifiedKFold、GroupKFold等详见sklearn官方文档。
拓展阅读