我使用面板数据：我随着时间的推移观察许多单位（例如人）；对于每个单元，我都有相同固定时间间隔的记录。
将数据拆分为训练集和测试集时，我们需要确保这两个集是不相交且顺序的，即训练集中的最新记录应位于测试集中最早的记录之前（例如，参见此博客）发帖）。
是否有面板数据交叉验证的标准 Python 实现？
我尝试过 Scikit-Learn 的 TimeSeriesSplit，它无法解释组，以及 GroupShuffleSplit，它无法解释数据的顺序性质，请参阅下面的代码。
import pandas as pd
import numpy as np
from sklearn.model_selection import GroupShuffleSplit, TimeSeriesSplit
# generate panel data
user = np.repeat(np.arange(10), 12)
time = np.tile(pd.date_range(start='2018-01-01', periods=12, freq='M'), 10)
data = (pd.DataFrame({'user': user, 'time': time})
.sort_values(['time', 'user'])
.reset_index(drop=True))
tscv = TimeSeriesSplit(n_splits=4)
for train_idx, test_idx in tscv.split(data):
train = data.iloc[train_idx]
test = data.iloc[test_idx]
train_end = train.time.max().date()
test_start = test.time.min().date()
print('TRAIN:', train_end, '\tTEST:', test_start, '\tSequential:', train_end < test_start, sep=' ')
输出：
TRAIN: 2018-03-31 TEST: 2018-03-31 Sequential: False
TRAIN: 2018-05-31 TEST: 2018-05-31 Sequential: False
TRAIN: 2018-08-31 TEST: 2018-08-31 Sequential: False
TRAIN: 2018-10-31 TEST: 2018-10-31 Sequential: False
因此，在这个示例中，我希望训练集和测试集仍然是连续的。
有许多相关的旧帖子，但没有（令人信服的）答案，请参见例如
此功能是在 scikit-learn 上请求的，我已为其添加了 PR 。 这项技术在最近的一些 Kaggle Notebooks 上得到了令人惊叹的结果。
gapfrom sklearn.model_selection._split import _BaseKFold, indexable, _num_samples
from sklearn.utils.validation import _deprecate_positional_args
# https://github.com/getgaurav2/scikit-learn/blob/d4a3af5cc9da3a76f0266932644b884c99724c57/sklearn/model_selection/_split.py#L2243
class GroupTimeSeriesSplit(_BaseKFold):
"""Time Series cross-validator variant with non-overlapping groups.
Provides train/test indices to split time series data samples
that are observed at fixed time intervals according to a
third-party provided group.
In each split, test indices must be higher than before, and thus shuffling
in cross validator is inappropriate.
This cross-validation object is a variation of :class:`KFold`.
In the kth split, it returns first k folds as train set and the
(k+1)th fold as test set.
The same group will not appear in two different folds (the number of
distinct groups has to be at least equal to the number of folds).
Note that unlike standard cross-validation methods, successive
training sets are supersets of those that come before them.
Read more in the :ref:`User Guide <cross_validation>`.
Parameters
----------
n_splits : int, default=5
Number of splits. Must be at least 2.
max_train_size : int, default=None
Maximum size for a single training set.
Examples
--------
>>> import numpy as np
>>> from sklearn.model_selection import GroupTimeSeriesSplit
>>> groups = np.array(['a', 'a', 'a', 'a', 'a', 'a',\
'b', 'b', 'b', 'b', 'b',\
'c', 'c', 'c', 'c',\
'd', 'd', 'd'])
>>> gtss = GroupTimeSeriesSplit(n_splits=3)
>>> for train_idx, test_idx in gtss.split(groups, groups=groups):
... print("TRAIN:", train_idx, "TEST:", test_idx)
... print("TRAIN GROUP:", groups[train_idx],\
"TEST GROUP:", groups[test_idx])
TRAIN: [0, 1, 2, 3, 4, 5] TEST: [6, 7, 8, 9, 10]
TRAIN GROUP: ['a' 'a' 'a' 'a' 'a' 'a']\
TEST GROUP: ['b' 'b' 'b' 'b' 'b']
TRAIN: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] TEST: [11, 12, 13, 14]
TRAIN GROUP: ['a' 'a' 'a' 'a' 'a' 'a' 'b' 'b' 'b' 'b' 'b']\
TEST GROUP: ['c' 'c' 'c' 'c']
TRAIN: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]\
TEST: [15, 16, 17]
TRAIN GROUP: ['a' 'a' 'a' 'a' 'a' 'a' 'b' 'b' 'b' 'b' 'b' 'c' 'c' 'c' 'c']\
TEST GROUP: ['d' 'd' 'd']
"""
@_deprecate_positional_args
def __init__(self,
n_splits=5,
*,
max_train_size=None
):
super().__init__(n_splits, shuffle=False, random_state=None)
self.max_train_size = max_train_size
def split(self, X, y=None, groups=None):
"""Generate indices to split data into training and test set.
Parameters
----------
X : array-like of shape (n_samples, n_features)
Training data, where n_samples is the number of samples
and n_features is the number of features.
y : array-like of shape (n_samples,)
Always ignored, exists for compatibility.
groups : array-like of shape (n_samples,)
Group labels for the samples used while splitting the dataset into
train/test set.
Yields
------
train : ndarray
The training set indices for that split.
test : ndarray
The testing set indices for that split.
"""
if groups is None:
raise ValueError(
"The 'groups' parameter should not be None")
X, y, groups = indexable(X, y, groups)
n_samples = _num_samples(X)
n_splits = self.n_splits
n_folds = n_splits + 1
group_dict = {}
u, ind = np.unique(groups, return_index=True)
unique_groups = u[np.argsort(ind)]
n_samples = _num_samples(X)
n_groups = _num_samples(unique_groups)
for idx in np.arange(n_samples):
if (groups[idx] in group_dict):
group_dict[groups[idx]].append(idx)
else:
group_dict[groups[idx]] = [idx]
if n_folds > n_groups:
raise ValueError(
("Cannot have number of folds={0} greater than"
" the number of groups={1}").format(n_folds,
n_groups))
group_test_size = n_groups // n_folds
group_test_starts = range(n_groups - n_splits * group_test_size,
n_groups, group_test_size)
for group_test_start in group_test_starts:
train_array = []
test_array = []
for train_group_idx in unique_groups[:group_test_start]:
train_array_tmp = group_dict[train_group_idx]
train_array = np.sort(np.unique(
np.concatenate((train_array,
train_array_tmp)),
axis=None), axis=None)
train_end = train_array.size
if self.max_train_size and self.max_train_size < train_end:
train_array = train_array[train_end -
self.max_train_size:train_end]
for test_group_idx in unique_groups[group_test_start:
group_test_start +
group_test_size]:
test_array_tmp = group_dict[test_group_idx]
test_array = np.sort(np.unique(
np.concatenate((test_array,
test_array_tmp)),
axis=None), axis=None)
yield [int(i) for i in train_array], [int(i) for i in test_array]
以 GridSearchCV 为例。代码修改自SO帖子here。
import xgboost as xgb
from sklearn.model_selection import GridSearchCV
import numpy as np
groups = np.array(['a', 'a', 'a', 'b', 'b', 'c'])
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])
model = xgb.XGBRegressor()
param_search = {'max_depth' : [3, 5]}
tscv = GroupTimeSeriesSplit(n_splits=2)
gsearch = GridSearchCV(estimator=model, cv=tscv,
param_grid=param_search)
gsearch.fit(X, y , groups=groups)
我最近完成了同样的任务，在找不到合适的解决方案后，我决定编写自己的类，这是 scikit-learn
TimeSeriesSplit这个想法基本上是按
datatimetimeTimeSeriesSplitimport numpy as np
from sklearn.utils import indexable
from sklearn.utils.validation import _num_samples
from sklearn.model_selection._split import _BaseKFold
class GroupTimeSeriesSplit(_BaseKFold):
"""
Time Series cross-validator for a variable number of observations within the time
unit. In the kth split, it returns first k folds as train set and the (k+1)th fold
as test set. Indices can be grouped so that they enter the CV fold together.
Parameters
----------
n_splits : int, default=5
Number of splits. Must be at least 2.
max_train_size : int, default=None
Maximum size for a single training set.
"""
def __init__(self, n_splits=5, *, max_train_size=None):
super().__init__(n_splits, shuffle=False, random_state=None)
self.max_train_size = max_train_size
def split(self, X, y=None, groups=None):
"""
Generate indices to split data into training and test set.
Parameters
----------
X : array-like of shape (n_samples, n_features)
Training data, where n_samples is the number of samples and n_features is
the number of features.
y : array-like of shape (n_samples,)
Always ignored, exists for compatibility.
groups : array-like of shape (n_samples,)
Group labels for the samples used while splitting the dataset into
train/test set.
Most often just a time feature.
Yields
-------
train : ndarray
The training set indices for that split.
test : ndarray
The testing set indices for that split.
"""
n_splits = self.n_splits
X, y, groups = indexable(X, y, groups)
n_samples = _num_samples(X)
n_folds = n_splits + 1
indices = np.arange(n_samples)
group_counts = np.unique(groups, return_counts=True)[1]
groups = np.split(indices, np.cumsum(group_counts)[:-1])
n_groups = _num_samples(groups)
if n_folds > n_groups:
raise ValueError(
("Cannot have number of folds ={0} greater"
" than the number of groups: {1}.").format(n_folds, n_groups))
test_size = (n_groups // n_folds)
test_starts = range(test_size + n_groups % n_folds,
n_groups, test_size)
for test_start in test_starts:
if self.max_train_size:
train_start = np.searchsorted(
np.cumsum(
group_counts[:test_start][::-1])[::-1] < self.max_train_size + 1,
True)
yield (np.concatenate(groups[train_start:test_start]),
np.concatenate(groups[test_start:test_start + test_size]))
else:
yield (np.concatenate(groups[:test_start]),
np.concatenate(groups[test_start:test_start + test_size]))
并将其应用到 OP 示例中，我们得到：
gtscv = GroupTimeSeriesSplit(n_splits=3)
for split_id, (train_id, val_id) in enumerate(gtscv.split(data, groups=data["time"])):
print("Split id: ", split_id, "\n")
print("Train id: ", train_id, "\n", "Validation id: ", val_id)
print("Train dates: ", data.loc[train_id, "time"].unique(), "\n", "Validation dates: ", data.loc[val_id, "time"].unique(), "\n")
Split id: 0
Train id: [ 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]
Validation id: [30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53
54 55 56 57 58 59]
Train dates: ['2018-01-31T00:00:00.000000000' '2018-02-28T00:00:00.000000000'
'2018-03-31T00:00:00.000000000']
Validation dates: ['2018-04-30T00:00:00.000000000' '2018-05-31T00:00:00.000000000'
'2018-06-30T00:00:00.000000000']
Split id: 1
Train id: [ 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 35 36 37 38 39 40 41 42 43 44 45 46 47
48 49 50 51 52 53 54 55 56 57 58 59]
Validation id: [60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
84 85 86 87 88 89]
Train dates: ['2018-01-31T00:00:00.000000000' '2018-02-28T00:00:00.000000000'
'2018-03-31T00:00:00.000000000' '2018-04-30T00:00:00.000000000'
'2018-05-31T00:00:00.000000000' '2018-06-30T00:00:00.000000000']
Validation dates: ['2018-07-31T00:00:00.000000000' '2018-08-31T00:00:00.000000000'
'2018-09-30T00:00:00.000000000']
Split id: 2
Train id: [ 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 35 36 37 38 39 40 41 42 43 44 45 46 47
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89]
Validation id: [ 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
108 109 110 111 112 113 114 115 116 117 118 119]
Train dates: ['2018-01-31T00:00:00.000000000' '2018-02-28T00:00:00.000000000'
'2018-03-31T00:00:00.000000000' '2018-04-30T00:00:00.000000000'
'2018-05-31T00:00:00.000000000' '2018-06-30T00:00:00.000000000'
'2018-07-31T00:00:00.000000000' '2018-08-31T00:00:00.000000000'
'2018-09-30T00:00:00.000000000']
Validation dates: ['2018-10-31T00:00:00.000000000' '2018-11-30T00:00:00.000000000'
'2018-12-31T00:00:00.000000000']