Adapted hv Block Cross-validation method

timecave.validation_methods.CV.AdaptedhvBlockCV(splits, ts, fs=1, h=1)

Bases: BaseSplitter

Implements the Adapted hv Block Cross-validation method.

This class implements the Adapted hv Block Cross-validation method. It is similar to the BlockCV class, but it does not support weight generation. Consequently, in order to implement a weighted version of this method, the user must implement their own derived class or compute the weights separately.

Parameters:

Name	Type	Description	Default
splits	int	The number of folds used to partition the data.	required
ts	ndarray | Series	Univariate time series.	required
fs	float | int	Sampling frequency (Hz).	1
h	int	Controls the amount of samples that will be removed from the training set. The h samples immediately following and preceding the validation set are not used for training.	1

Attributes:

Name	Type	Description
n_splits	int	The number of splits.
sampling_freq	int | float	The series' sampling frequency (Hz).

Methods:

Name	Description
split	Split the time series into training and validation sets.
info	Provide additional information on the validation method.
statistics	Compute relevant statistics for both training and validation sets.
plot	Plot the partitioned time series.

Raises:

Type	Description
TypeError	If h is not an integer.
ValueError	If h is smaller than or equal to zero.
ValueError	If h is larger than the number of samples in a fold.

See also

Block CV: The original Block CV method, where no training samples are removed.

Notes

The Adapted hv Block Cross-validation method splits the data into \(N\) different folds. Then, in every iteration \(i\), the model is validated on data from the \(i^{th}\) folds and trained on data from the remaining folds. There is, however, one subtle difference from the original Block Cross-validation method: the \(h\) training samples that lie closest to the validation set are removed, thereby reducing the correlation between the training set and the validation set. The average error on the validation sets is then taken as the estimate of the model's true error. This method does not preserve the temporal order of the observations.

This method was proposed by Cerqueira et al. [1].

References

1

Vitor Cerqueira, Luis Torgo, and Igor Mozetiˇc. Evaluating time series forecasting models: An empirical study on performance estimation methods. Machine Learning, 109(11):1997–2028, 2020.

Source code in timecave/validation_methods/CV.py

def __init__(
    self,
    splits: int,
    ts: np.ndarray | pd.Series,
    fs: float | int = 1,
    h: int = 1,
) -> None:

    super().__init__(splits, ts, fs)
    self._check_h(h);
    self._h = h;
    self._splitting_ind = self._split_ind()

info()

Provide some basic information on the training and validation sets.

This method displays the number of splits and the fold size.

Examples:

>>> import numpy as np
>>> from timecave.validation_methods.CV import AdaptedhvBlockCV
>>> ts = np.ones(10);
>>> splitter = AdaptedhvBlockCV(5, ts);
>>> splitter.info();
Adapted hv Block CV method
---------------
Time series size: 10 samples
Number of splits: 5
Fold size: 1 to 2 samples (10.0 to 20.0 %)

Source code in timecave/validation_methods/CV.py

def info(self) -> None:
    """
    Provide some basic information on the training and validation sets.

    This method displays the number of splits and the fold size.

    Examples
    --------
    >>> import numpy as np
    >>> from timecave.validation_methods.CV import AdaptedhvBlockCV
    >>> ts = np.ones(10);
    >>> splitter = AdaptedhvBlockCV(5, ts);
    >>> splitter.info();
    Adapted hv Block CV method
    ---------------
    Time series size: 10 samples
    Number of splits: 5
    Fold size: 1 to 2 samples (10.0 to 20.0 %)
    """

    min_fold_size = int(np.floor(self._n_samples / self.n_splits)) - self._h
    max_fold_size = int(np.floor(self._n_samples / self.n_splits))

    remainder = self._n_samples % self.n_splits

    if remainder != 0:

        max_fold_size += 1

    min_fold_size_pct = np.round(min_fold_size / self._n_samples * 100, 2)
    max_fold_size_pct = np.round(max_fold_size / self._n_samples * 100, 2)

    print("Adapted hv Block CV method")
    print("---------------")
    print(f"Time series size: {self._n_samples} samples")
    print(f"Number of splits: {self.n_splits}")
    print(
        f"Fold size: {min_fold_size} to {max_fold_size} samples ({min_fold_size_pct} to {max_fold_size_pct} %)"
    )

    return

plot(height, width)

Plot the partitioned time series.

This method allows the user to plot the partitioned time series. The training and validation sets are plotted using different colours.

Parameters:

Name	Type	Description	Default
height	int	The figure's height.	required
width	int	The figure's width.	required

Examples:

>>> import numpy as np
>>> from timecave.validation_methods.CV import AdaptedhvBlockCV
>>> ts = np.ones(100);
>>> splitter = AdaptedhvBlockCV(5, ts, h=5);
>>> splitter.plot(10, 10);

Source code in timecave/validation_methods/CV.py

def plot(self, height: int, width: int) -> None:
    """
    Plot the partitioned time series.

    This method allows the user to plot the partitioned time series. The training and validation sets are plotted using different colours.

    Parameters
    ----------
    height : int
        The figure's height.

    width : int
        The figure's width.

    Examples
    --------
    >>> import numpy as np
    >>> from timecave.validation_methods.CV import AdaptedhvBlockCV
    >>> ts = np.ones(100);
    >>> splitter = AdaptedhvBlockCV(5, ts, h=5);
    >>> splitter.plot(10, 10);

    ![block_plot](../../../images/AdaptedHV_plot.png)
    """

    fig, axs = plt.subplots(self.n_splits, 1, sharex=True)
    fig.set_figheight(height)
    fig.set_figwidth(width)
    fig.supxlabel("Samples")
    fig.supylabel("Time Series")
    fig.suptitle("Adapted hv Block CV method")

    for it, (training, validation, w) in enumerate(self.split()):

        axs[it].scatter(training, self._series[training], label="Training set")
        axs[it].scatter(
            validation, self._series[validation], label="Validation set"
        )
        axs[it].set_title("Fold: {}".format(it + 1))
        axs[it].set_ylim([self._series.min() - 1, self._series.max() + 1])
        axs[it].set_xlim([- 1, self._n_samples + 1])
        axs[it].legend()

    plt.show()

    return

split()

Split the time series into training and validation sets.

This method splits the series' indices into disjoint sets containing the training and validation indices. At every iteration, an array of training indices and another one containing the validation indices are generated. Note that this method is a generator. To access the indices, use the next() method or a for loop.

Yields:

Type	Description
ndarray	Array of training indices.
ndarray	Array of validation indices.
float	Weight assigned the error estimate.

Examples:

>>> import numpy as np
>>> from timecave.validation_methods.CV import AdaptedhvBlockCV
>>> ts = np.ones(10);
>>> splitter = AdaptedhvBlockCV(5, ts); # Split the data into 5 different folds
>>> for ind, (train, val, _) in enumerate(splitter.split()):
... 
...     print(f"Iteration {ind+1}");
...     print(f"Training set indices: {train}");
...     print(f"Validation set indices: {val}");
Iteration 1
Training set indices: [3 4 5 6 7 8 9]
Validation set indices: [0 1]
Iteration 2
Training set indices: [0 5 6 7 8 9]
Validation set indices: [2 3]
Iteration 3
Training set indices: [0 1 2 7 8 9]
Validation set indices: [4 5]
Iteration 4
Training set indices: [0 1 2 3 4 9]
Validation set indices: [6 7]
Iteration 5
Training set indices: [0 1 2 3 4 5 6]
Validation set indices: [8 9]

If the number of samples is not divisible by the number of folds, the first folds will contain more samples:

>>> ts2 = np.ones(17);
>>> splitter = AdaptedhvBlockCV(5, ts2, h=2);
>>> for ind, (train, val, _) in enumerate(splitter.split()):
... 
...     print(f"Iteration {ind+1}");
...     print(f"Training set indices: {train}");
...     print(f"Validation set indices: {val}");
Iteration 1
Training set indices: [ 6  7  8  9 10 11 12 13 14 15 16]
Validation set indices: [0 1 2 3]
Iteration 2
Training set indices: [ 0  1 10 11 12 13 14 15 16]
Validation set indices: [4 5 6 7]
Iteration 3
Training set indices: [ 0  1  2  3  4  5 13 14 15 16]
Validation set indices: [ 8  9 10]
Iteration 4
Training set indices: [ 0  1  2  3  4  5  6  7  8 16]
Validation set indices: [11 12 13]
Iteration 5
Training set indices: [ 0  1  2  3  4  5  6  7  8  9 10 11]
Validation set indices: [14 15 16]

Source code in timecave/validation_methods/CV.py

def split(self) -> Generator[tuple[np.ndarray, np.ndarray, float], None, None]:
    """
    Split the time series into training and validation sets.

    This method splits the series' indices into disjoint sets containing the training and validation indices.
    At every iteration, an array of training indices and another one containing the validation indices are generated.
    Note that this method is a generator. To access the indices, use the `next()` method or a `for` loop.

    Yields
    ------
    np.ndarray
        Array of training indices.

    np.ndarray
        Array of validation indices.

    float
        Weight assigned the error estimate.

    Examples
    --------
    >>> import numpy as np
    >>> from timecave.validation_methods.CV import AdaptedhvBlockCV
    >>> ts = np.ones(10);
    >>> splitter = AdaptedhvBlockCV(5, ts); # Split the data into 5 different folds
    >>> for ind, (train, val, _) in enumerate(splitter.split()):
    ... 
    ...     print(f"Iteration {ind+1}");
    ...     print(f"Training set indices: {train}");
    ...     print(f"Validation set indices: {val}");
    Iteration 1
    Training set indices: [3 4 5 6 7 8 9]
    Validation set indices: [0 1]
    Iteration 2
    Training set indices: [0 5 6 7 8 9]
    Validation set indices: [2 3]
    Iteration 3
    Training set indices: [0 1 2 7 8 9]
    Validation set indices: [4 5]
    Iteration 4
    Training set indices: [0 1 2 3 4 9]
    Validation set indices: [6 7]
    Iteration 5
    Training set indices: [0 1 2 3 4 5 6]
    Validation set indices: [8 9]

    If the number of samples is not divisible by the number of folds, the first folds will contain more samples:

    >>> ts2 = np.ones(17);
    >>> splitter = AdaptedhvBlockCV(5, ts2, h=2);
    >>> for ind, (train, val, _) in enumerate(splitter.split()):
    ... 
    ...     print(f"Iteration {ind+1}");
    ...     print(f"Training set indices: {train}");
    ...     print(f"Validation set indices: {val}");
    Iteration 1
    Training set indices: [ 6  7  8  9 10 11 12 13 14 15 16]
    Validation set indices: [0 1 2 3]
    Iteration 2
    Training set indices: [ 0  1 10 11 12 13 14 15 16]
    Validation set indices: [4 5 6 7]
    Iteration 3
    Training set indices: [ 0  1  2  3  4  5 13 14 15 16]
    Validation set indices: [ 8  9 10]
    Iteration 4
    Training set indices: [ 0  1  2  3  4  5  6  7  8 16]
    Validation set indices: [11 12 13]
    Iteration 5
    Training set indices: [ 0  1  2  3  4  5  6  7  8  9 10 11]
    Validation set indices: [14 15 16]
    """

    for i, (ind) in enumerate(self._splitting_ind[:-1]):

        next_ind = self._splitting_ind[i + 1]

        validation = self._indices[ind:next_ind]
        h_ind = self._indices[
            np.fmax(ind - self._h, 0) : np.fmin(
                next_ind + self._h, self._n_samples
            )
        ]
        train = np.array([el for el in self._indices if el not in h_ind])

        yield (train, validation, 1.0)

statistics()

Compute relevant statistics for both training and validation sets.

This method computes relevant time series features, such as mean, strength-of-trend, etc. for both the whole time series, the training set and the validation set. It can and should be used to ensure that the characteristics of both the training and validation sets are, statistically speaking, similar to those of the time series one wishes to forecast. If this is not the case, using the validation method will most likely lead to a poor assessment of the model's performance.

Returns:

Type	Description
DataFrame	Relevant features for the entire time series.
DataFrame	Relevant features for the training set.
DataFrame	Relevant features for the validation set.

Raises:

Type	Description
ValueError	If the time series is composed of less than three samples.
ValueError	If the folds comprise less than two samples.

Examples:

>>> import numpy as np
>>> from timecave.validation_methods.CV import AdaptedhvBlockCV
>>> ts = np.hstack((np.ones(5), np.zeros(5)));
>>> splitter = AdaptedhvBlockCV(5, ts);
>>> ts_stats, training_stats, validation_stats = splitter.statistics();
>>> ts_stats
   Mean  Median  Min  Max  Variance  P2P_amplitude  Trend_slope  Spectral_centroid  Spectral_rolloff  Spectral_entropy  Strength_of_trend  Mean_crossing_rate  Median_crossing_rate
0   0.5     0.5  0.0  1.0      0.25            1.0    -0.151515           0.114058               0.5           0.38717            1.59099            0.111111              0.111111
>>> training_stats
       Mean  Median  Min  Max  Variance  P2P_amplitude  Trend_slope  Spectral_centroid  Spectral_rolloff  Spectral_entropy  Strength_of_trend  Mean_crossing_rate  Median_crossing_rate
0  0.285714     0.0  0.0  1.0  0.204082            1.0    -0.178571           0.146421          0.428571          0.702232           1.212183            0.166667              0.166667
0  0.166667     0.0  0.0  1.0  0.138889            1.0    -0.142857           0.250000          0.500000          1.000000           0.931695            0.200000              0.200000
0  0.500000     0.5  0.0  1.0  0.250000            1.0    -0.257143           0.138889          0.500000          0.455486           1.250000            0.200000              0.200000
0  0.833333     1.0  0.0  1.0  0.138889            1.0    -0.142857           0.125000          0.500000          0.792481           0.931695            0.200000              0.200000
0  0.714286     1.0  0.0  1.0  0.204082            1.0    -0.178571           0.094706          0.428571          0.556506           1.212183            0.166667              0.166667
>>> validation_stats
   Mean  Median  Min  Max  Variance  P2P_amplitude   Trend_slope  Spectral_centroid  Spectral_rolloff  Spectral_entropy  Strength_of_trend  Mean_crossing_rate  Median_crossing_rate
0   1.0     1.0  1.0  1.0      0.00            0.0 -7.850462e-17               0.00               0.0               0.0                inf                 0.0                   0.0
0   1.0     1.0  1.0  1.0      0.00            0.0 -7.850462e-17               0.00               0.0               0.0                inf                 0.0                   0.0
0   0.5     0.5  0.0  1.0      0.25            1.0 -1.000000e+00               0.25               0.5               0.0                inf                 1.0                   1.0
0   0.0     0.0  0.0  0.0      0.00            0.0  0.000000e+00               0.00               0.0               0.0                inf                 0.0                   0.0
0   0.0     0.0  0.0  0.0      0.00            0.0  0.000000e+00               0.00               0.0               0.0                inf                 0.0                   0.0

Source code in timecave/validation_methods/CV.py

def statistics(self) -> tuple[pd.DataFrame]:
    """
    Compute relevant statistics for both training and validation sets.

    This method computes relevant time series features, such as mean, strength-of-trend, etc. for both the whole time series, the training set and the validation set.
    It can and should be used to ensure that the characteristics of both the training and validation sets are, statistically speaking, similar to those of the time series one wishes to forecast.
    If this is not the case, using the validation method will most likely lead to a poor assessment of the model's performance.

    Returns
    -------
    pd.DataFrame
        Relevant features for the entire time series.

    pd.DataFrame
        Relevant features for the training set.

    pd.DataFrame
        Relevant features for the validation set.

    Raises
    ------
    ValueError
        If the time series is composed of less than three samples.

    ValueError
        If the folds comprise less than two samples.

    Examples
    --------
    >>> import numpy as np
    >>> from timecave.validation_methods.CV import AdaptedhvBlockCV
    >>> ts = np.hstack((np.ones(5), np.zeros(5)));
    >>> splitter = AdaptedhvBlockCV(5, ts);
    >>> ts_stats, training_stats, validation_stats = splitter.statistics();
    >>> ts_stats
       Mean  Median  Min  Max  Variance  P2P_amplitude  Trend_slope  Spectral_centroid  Spectral_rolloff  Spectral_entropy  Strength_of_trend  Mean_crossing_rate  Median_crossing_rate
    0   0.5     0.5  0.0  1.0      0.25            1.0    -0.151515           0.114058               0.5           0.38717            1.59099            0.111111              0.111111
    >>> training_stats
           Mean  Median  Min  Max  Variance  P2P_amplitude  Trend_slope  Spectral_centroid  Spectral_rolloff  Spectral_entropy  Strength_of_trend  Mean_crossing_rate  Median_crossing_rate
    0  0.285714     0.0  0.0  1.0  0.204082            1.0    -0.178571           0.146421          0.428571          0.702232           1.212183            0.166667              0.166667
    0  0.166667     0.0  0.0  1.0  0.138889            1.0    -0.142857           0.250000          0.500000          1.000000           0.931695            0.200000              0.200000
    0  0.500000     0.5  0.0  1.0  0.250000            1.0    -0.257143           0.138889          0.500000          0.455486           1.250000            0.200000              0.200000
    0  0.833333     1.0  0.0  1.0  0.138889            1.0    -0.142857           0.125000          0.500000          0.792481           0.931695            0.200000              0.200000
    0  0.714286     1.0  0.0  1.0  0.204082            1.0    -0.178571           0.094706          0.428571          0.556506           1.212183            0.166667              0.166667
    >>> validation_stats
       Mean  Median  Min  Max  Variance  P2P_amplitude   Trend_slope  Spectral_centroid  Spectral_rolloff  Spectral_entropy  Strength_of_trend  Mean_crossing_rate  Median_crossing_rate
    0   1.0     1.0  1.0  1.0      0.00            0.0 -7.850462e-17               0.00               0.0               0.0                inf                 0.0                   0.0
    0   1.0     1.0  1.0  1.0      0.00            0.0 -7.850462e-17               0.00               0.0               0.0                inf                 0.0                   0.0
    0   0.5     0.5  0.0  1.0      0.25            1.0 -1.000000e+00               0.25               0.5               0.0                inf                 1.0                   1.0
    0   0.0     0.0  0.0  0.0      0.00            0.0  0.000000e+00               0.00               0.0               0.0                inf                 0.0                   0.0
    0   0.0     0.0  0.0  0.0      0.00            0.0  0.000000e+00               0.00               0.0               0.0                inf                 0.0                   0.0
    """

    if self._n_samples <= 2:

        raise ValueError(
            "Basic statistics can only be computed if the time series comprises more than two samples."
        )

    if int(np.round(self._n_samples / self.n_splits)) < 2:

        raise ValueError(
            "The folds are too small to compute most meaningful features."
        )

    full_features = get_features(self._series, self.sampling_freq)
    training_stats = []
    validation_stats = []

    for (training, validation, _) in self.split():

        training_feat = get_features(self._series[training], self.sampling_freq)
        training_stats.append(training_feat)

        validation_feat = get_features(self._series[validation], self.sampling_freq)
        validation_stats.append(validation_feat)

    training_features = pd.concat(training_stats)
    validation_features = pd.concat(validation_stats)

    return (full_features, training_features, validation_features)