wildboar.model_selection.outlier#

Module Contents#

Classes#

RepeatedOutlierSplit

Repeated random outlier cross-validator

Functions#

train_test_split(x, y, normal_class[, test_size, ...])

Training and testing split from classification dataset

threshold_score(y_true, score, score_f)

Compute the performance of using the i:th score
class wildboar.model_selection.outlier.RepeatedOutlierSplit(n_splits=None, *, test_size=0.2, n_outlier=0.05, shuffle=True, random_state=None)#

Repeated random outlier cross-validator

Yields indicies that split the dataset into training and test sets.

Note

Contrary to other cross-validation strategies, the random outlier cross-validator does not ensure that all folds will be different. Instead, the inlier samples are shuffled and new outlier samples are inserted in the training and test sets repeatedly.

get_n_splits(X, y, groups=None)#

Returns the number of splitting iterations in the cross-validator :param X: The samples :type X: object :param y: The labels :type y: object :param groups: Always ignored, exists for compatibility. :type groups: object

Returns:

n_splits – Returns the number of splitting iterations in the cross-validator.

Return type:

int

split(x, y, groups=None)#

Return training and test indicies

Parameters:

  • x (object) – Always ignored, exists for compatibility.

  • y (object) – The labels

  • groups (object, optional) – Always ignored, exists for compatibility.

Yields:

train_idx, test_idx (ndarray) – The training and test indicies

__repr__()#

Return repr(self).

wildboar.model_selection.outlier.train_test_split(x, y, normal_class, test_size=0.2, anomalies_train_size=0.05, random_state=None)#

Training and testing split from classification dataset

Parameters:

  • x (array-like of shape (n_samples, n_timestep) or (n_samples, n_dim, n_timestep)) – Input data samples

  • y (array-like of shape (n_samples,)) – Input class label

  • normal_class (int) – Class label that should be considered as the normal class

  • test_size (float) – Size of the test set

  • anomalies_train_size (float) – Contamination of anomalies in the training dataset

  • random_state (int or RandomState) – Psudo random state used for stable results.

Returns:

  • x_train (array-like) – Training samples

  • x_test (array-like) – Test samples

  • y_train (array-like) – Training labels (either 1 or -1, where 1 denotes normal and -1 anomalous)

  • y_test (array-like) – Test labels (either 1 or -1, where 1 denotes normal and -1 anomalous)

Examples

>>> from wildboar.datasets import load_two_lead_ecg
>>> x, y = load_two_lead_ecg()
>>> x_train, x_test, y_train, y_test = train_test_split(x, y, 1, test_size=0.2, anomalies_train_size=0.05)
wildboar.model_selection.outlier.threshold_score(y_true, score, score_f)#

Compute the performance of using the i:th score

The scores are typically computed using an outlier detection algorithm

Parameters:

  • y_true (array-like) – The true labels

  • score (array-like) – The scores

  • score_f (callable) – Function for estimating the performance of the i:th scoring

Returns:

score – performance for each score as threshold

Return type:

ndarray

See also

wildboar.ensemble.IsolationShapeletForest

an isolation forest for time series

Examples

Setting the offset that maximizes balanced accuracy of a shapelet isolation forest

>>> from wildboar.ensemble import IsolationShapeletForest
>>> from wildboar.datasets import load_two_lead_ecg
>>> from sklearn.metrics import balanced_accuracy_score
>>> x, y = load_two_lead_ecg()
>>> x_train, x_test, y_train, y_test = train_test_split(x, y, 1, test_size=0.2, anomalies_train_size=0.05)
>>> f = IsolationShapeletForest()
>>> f.fit(x_train)
>>> scores = f.score_samples(x_train)
>>> perf = threshold_score(y_train, scores, balanced_accuracy_score)
>>> f.offset_ = score[np.argmax(perf)]