Sktime中的窗口拆分

Sktime中的窗口拆分

在此notebook中，我们将描述 sktime.forecasting.model_selection模块下的窗口拆分。窗口拆分可以和 ForecastingGridSearchCV一起来做模型选择（参考forecasting notebook）。

**注意：**需要强调的是，在做时间序列的交叉验证时，不能随机打散（shuffle）数据集，否则会导致信息泄漏（leaking information）。

参考：

• Cross-validation: evaluating estimator performance
• Cross-validation for time series

准备工作

import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
from matplotlib.ticker import MaxNLocator

from sktime.datasets import load_airline
from sktime.forecasting.base import ForecastingHorizon

# 4种窗口拆分方式
from sktime.forecasting.model_selection import (
    CutoffSplitter,
    SingleWindowSplitter,
    SlidingWindowSplitter,
    temporal_train_test_split,
)

# 时序绘图
from sktime.utils.plotting import plot_series

# 绘图参数
sns.set_style("whitegrid")
plt.rcParams["figure.figsize"] = [12, 6]
plt.rcParams["figure.dpi"] = 100

数据

我们将使用航空公司数据集（Box-Jenkins，单变量）的一小部分，该数据集包含1949-1960年间每个月国际航空公司乘客的数量。

y = load_airline().iloc[:30]
y.head()

# Period
# 1949-01    112.0
# 1949-02    118.0
# 1949-03    132.0
# 1949-04    129.0
# 1949-05    121.0
# Freq: M, Name: Number of airline passengers, dtype: float64

fig, ax = plot_series(y)

# 采用 forecasting horizon
fh = ForecastingHorizon([1, 2, 3, 4, 5])
y_train, y_test = temporal_train_test_split(y, fh=fh)
plot_series(y_train, y_test, labels=["y_train", "y_test"]);

窗口拆分

现在我们描述不同的拆分方法

1.temporal_train_test_split

把数据集拆分为 traininig 和test 。你可以采用以下两种方式：

• 设定training 或者 test 的 size
• 设定预测期数（forecasting horizon）

# 设定 test set size
y_train, y_test = temporal_train_test_split(y=y, test_size=0.25)
fig, ax = plot_series(y_train, y_test, labels=["y_train", "y_test"])

# 设定 forecasting horizon
fh = ForecastingHorizon([1, 2, 3, 4, 5])
y_train, y_test = temporal_train_test_split(y, fh=fh)
plot_series(y_train, y_test, labels=["y_train", "y_test"]);

2.SingleWindowSplitter

这个类一次性把时间序列拆分为training和test两个窗口 。这和temporal_train_test_split有些类似。

首先，先定义好折数（fold）中的参数：

# 定义拆分窗口的参数
window_length = 10
fh = ForecastingHorizon([1, 2, 3])
fh_length = len(fh)

cv = SingleWindowSplitter(window_length=window_length, fh=fh)

n_splits = cv.get_n_splits(y)
print(f"Number of Folds = {n_splits}")

# Number of Folds = 1

我们来绘图看看是怎么产生这1个fold的。首先定义以下一些帮助函数

def get_folds_arrays(y, cv):
    """Store folds as arrays."""
    n_splits = cv.get_n_splits(y)

    windows = np.empty((n_splits, window_length), dtype=np.int)
    fhs = np.empty((n_splits, fh_length), dtype=np.int)
    for i, (w, f) in enumerate(cv.split(y)):
        windows[i] = w
        fhs[i] = f
    return windows, fhs


def get_y(length, split):
    """Creates a constant level vector based on the split."""
    return np.ones(length) * split

windows, fhs = get_folds_arrays(y, cv)

window_color, fh_color = sns.color_palette("colorblind")[:2]

fig, ax = plt.subplots()
for i in range(n_splits):
    ax.plot(np.arange(len(y)), get_y(len(y), i), marker="o", c="lightgray")
    ax.plot(
        windows[i], get_y(window_length, i), marker="o", c=window_color, label="Window"
    )
    ax.plot(
        fhs[i], get_y(fh_length, i), marker="o", c=fh_color, label="Forecasting horizon"
    )
ax.invert_yaxis()
ax.yaxis.set_major_locator(MaxNLocator(integer=True))
ax.set(
    title="SingleWindowSplitter Fold",
    ylabel="Window number",
    xlabel="Time",
    xticklabels=y.index,
    ylim=(-1, 1),
)
# remove duplicate labels/handles
handles, labels = [(leg[:2]) for leg in ax.get_legend_handles_labels()]
ax.legend(handles, labels);

3.SlidingWindowSplitter

This splitter generates folds which move with time. The length of the training and test sets for each fold remains constant.

cv = SlidingWindowSplitter(window_length=window_length, fh=fh, start_with_window=True)

n_splits = cv.get_n_splits(y)
print(f"Number of Folds = {n_splits}")

windows, fhs = get_folds_arrays(y, cv)

fig, ax = plt.subplots()
for i in range(n_splits):
    ax.plot(np.arange(len(y)), get_y(len(y), i), marker="o", c="lightgray")
    ax.plot(
        windows[i], get_y(window_length, i), marker="o", c=window_color, label="Window"
    )
    ax.plot(
        fhs[i], get_y(fh_length, i), marker="o", c=fh_color, label="Forecasting horizon"
    )
ax.invert_yaxis()
ax.yaxis.set_major_locator(MaxNLocator(integer=True))
ax.set(
    title="SlidingWindowSplitter Folds",
    ylabel="Window number",
    xlabel="Time",
    xticklabels=y.index,
)
# remove duplicate labels/handles
handles, labels = [(leg[:2]) for leg in ax.get_legend_handles_labels()]
ax.legend(handles, labels);

4.CutoffSplitter

With this splitter we can manually select the cutoff points.

# Specify cutoff points (by array index).
cutoffs = np.array([10, 15, 20, 25])

cv = CutoffSplitter(cutoffs=cutoffs, window_length=window_length, fh=fh)

n_splits = cv.get_n_splits(y)
print(f"Number of Folds = {n_splits}")

windows, fhs = get_folds_arrays(y, cv)

fig, ax = plt.subplots()
for i in range(n_splits):
    ax.plot(np.arange(len(y)), get_y(len(y), i), marker="o", c="lightgray")
    ax.plot(
        windows[i], get_y(window_length, i), marker="o", c=window_color, label="Window"
    )
    ax.plot(
        fhs[i], get_y(fh_length, i), marker="o", c=fh_color, label="Forecasting horizon"
    )
ax.invert_yaxis()
ax.yaxis.set_major_locator(MaxNLocator(integer=True))
ax.set(
    title="CutoffSplitter Folds",
    ylabel="Window number",
    xlabel="Time",
    xticklabels=y.index,
)
# remove duplicate labels/handles
handles, labels = [(leg[:2]) for leg in ax.get_legend_handles_labels()]
ax.legend(handles, labels);