在 GridSearchCV 中明确指定 test/train 集合

Question

我对 sklearn 的 GridSearchCV 的 cv 参数有疑问。

我正在处理具有时间成分的数据，因此我认为 KFold 交叉验证中的随机改组似乎不明智。

相反，我想在 GridSearchCV 中明确指定训练、验证和测试数据的截止值。我可以这样做吗？

为了更好地阐明问题，以下是我手动解决该问题的方法。

import numpy as np
import pandas as pd
from sklearn.linear_model import Ridge
np.random.seed(444)

index = pd.date_range('2014', periods=60, freq='M')
X, y = make_regression(n_samples=60, n_features=3, random_state=444, noise=90.)
X = pd.DataFrame(X, index=index, columns=list('abc'))
y = pd.Series(y, index=index, name='y')

# Train on the first 30 samples, validate on the next 10, test on
#     the final 10.
X_train, X_val, X_test = np.array_split(X, [35, 50])
y_train, y_val, y_test = np.array_split(y, [35, 50])

param_grid = {'alpha': np.linspace(0, 1, 11)}
model = None
best_param_ = None
best_score_ = -np.inf

# Manual implementation
for alpha in param_grid['alpha']:
    ridge = Ridge(random_state=444, alpha=alpha).fit(X_train, y_train)
    score = ridge.score(X_val, y_val)
    if score > best_score_:
        best_score_ = score
        best_param_ = alpha
        model = ridge

print('Optimal alpha parameter: {:0.2f}'.format(best_param_))
print('Best score (on validation data): {:0.2f}'.format(best_score_))
print('Test set score: {:.2f}'.format(model.score(X_test, y_test)))
# Optimal alpha parameter: 1.00
# Best score (on validation data): 0.64
# Test set score: 0.22

这里的流程是：

对于X和Y，我都想要训练集、验证集和测试集。训练集是时间序列中的前 35 个样本。验证集是接下来的 15 个样本。测试集是最后的10.
训练集和验证集用于确定 Ridge 回归中的最佳 alpha 参数。在这里我测试 alphas of (0.0, 0.1, ..., 0.9, 1.0).
测试集作为未见数据用于 "actual" 测试。

总之...我似乎想做这样的事情，但不确定要传递给 cv 的内容：

from sklearn.model_selection import GridSearchCV
grid_search = GridSearchCV(Ridge(random_state=444), param_grid, cv= ???)
grid_search.fit(...?)

我无法解释的文档指定：

cv : int, cross-validation generator or an iterable, optional

Determines the cross-validation splitting strategy. Possible inputs for cv are:

None, to use the default 3-fold cross validation,

integer, to specify the number of folds in a (Stratified)KFold,

An object to be used as a cross-validation generator.

An iterable yielding train, test splits.

For integer/None inputs, if the estimator is a classifier and y is either binary or multiclass, StratifiedKFold is used. In all other cases, KFold is used.

Answer 1

正如@MaxU 所说，最好让 GridSearchCV 处理拆分，但是如果你想按照你在问题中设置的那样强制执行拆分，那么你可以使用 PredefinedSplit 来做到这一点东西.

因此您需要对代码进行以下更改。

# Here X_test, y_test is the untouched data
# Validation data (X_val, y_val) is currently inside X_train, which will be split using PredefinedSplit inside GridSearchCV
X_train, X_test = np.array_split(X, [50])
y_train, y_test = np.array_split(y, [50])


# The indices which have the value -1 will be kept in train.
train_indices = np.full((35,), -1, dtype=int)

# The indices which have zero or positive values, will be kept in test
test_indices = np.full((15,), 0, dtype=int)
test_fold = np.append(train_indices, test_indices)

print(test_fold)
# OUTPUT: 
array([-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
       -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
       -1,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0])

from sklearn.model_selection import PredefinedSplit
ps = PredefinedSplit(test_fold)

# Check how many splits will be done, based on test_fold
ps.get_n_splits()
# OUTPUT: 1

for train_index, test_index in ps.split():
    print("TRAIN:", train_index, "TEST:", test_index)

# OUTPUT: 
('TRAIN:', array([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16,
   17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33,
   34]), 
 'TEST:', array([35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49]))


# And now, send this `ps` to cv param in GridSearchCV
from sklearn.model_selection import GridSearchCV
grid_search = GridSearchCV(Ridge(random_state=444), param_grid, cv=ps)

# Here, send the X_train and y_train
grid_search.fit(X_train, y_train)

发送到 fit() 的 X_train、y_train 将使用我们定义的拆分拆分为训练和测试（在您的情况下为 val），因此，Ridge 将被训练基于索引 [0:35] 的原始数据并在 [35:50].

上进行测试

希望这能清除工作。

Answer 2

你试过了吗TimeSeriesSplit？

它是明确用于拆分时间序列数据的。

tscv = TimeSeriesSplit(n_splits=3)
grid_search = GridSearchCV(clf, param_grid, cv=tscv.split(X))

Answer 3

在时间序列数据中，Kfold 不是正确的方法，因为 kfold cv 会打乱你的数据，你会失去序列中的模式。这是一个方法

import xgboost as xgb
from sklearn.model_selection import TimeSeriesSplit, GridSearchCV
import numpy as np
X = np.array([[4, 5, 6, 1, 0, 2], [3.1, 3.5, 1.0, 2.1, 8.3, 1.1]]).T
y = np.array([1, 6, 7, 1, 2, 3])
tscv = TimeSeriesSplit(n_splits=2)

model = xgb.XGBRegressor()
param_search = {'max_depth' : [3, 5]}

my_cv = TimeSeriesSplit(n_splits=2).split(X)
gsearch = GridSearchCV(estimator=model, cv=my_cv,
                        param_grid=param_search)
gsearch.fit(X, y)

参考 -

在 GridSearchCV 中明确指定 test/train 集合

Explicitly specifying test/train sets in GridSearchCV

python

scikit-learn

grid-search