from sklearn import metrics
import numpy as np
from deepod.metrics.affiliation.generics import convert_vector_to_events
from deepod.metrics.vus.metrics import get_range_vus_roc
from deepod.metrics.affiliation.metrics import pr_from_events
from deepod.metrics._tsad_adjustment import point_adjustment
[docs]def auc_roc(y_true, y_score):
"""
Calculates the area under the Receiver Operating Characteristic (ROC) curve.
Args:
y_true (np.array, required):
True binary labels. 0 indicates a normal timestamp, and 1 indicates an anomaly.
y_score (np.array, required):
Predicted anomaly scores. A higher score indicates a higher likelihood of being an anomaly.
Returns:
float:
The score of the area under the ROC curve.
"""
return metrics.roc_auc_score(y_true, y_score)
[docs]def auc_pr(y_true, y_score):
"""
Calculates the area under the Precision-Recall (PR) curve.
Args:
y_true (np.array, required):
True binary labels. 0 indicates a normal timestamp, and 1 indicates an anomaly.
y_score (np.array, required):
Predicted anomaly scores. A higher score indicates a higher likelihood of being an anomaly.
Returns:
float:
The score of the area under the PR curve.
"""
return metrics.average_precision_score(y_true, y_score)
[docs]def tabular_metrics(y_true, y_score):
"""
Calculates evaluation metrics for tabular anomaly detection.
Args:
y_true (np.array, required):
Data label, 0 indicates normal timestamp, and 1 is anomaly.
y_score (np.array, required):
Predicted anomaly scores, higher score indicates higher likelihoods to be anomaly.
Returns:
tuple: A tuple containing:
- auc_roc (float):
The score of area under the ROC curve.
- auc_pr (float):
The score of area under the precision-recall curve.
- f1 (float):
The score of F1-score.
"""
# F1@k, using real percentage to calculate F1-score
ratio = 100.0 * len(np.where(y_true == 0)[0]) / len(y_true)
thresh = np.percentile(y_score, ratio)
y_pred = (y_score >= thresh).astype(int)
y_true = y_true.astype(int)
p, r, f1, support = metrics.precision_recall_fscore_support(y_true, y_pred, average='binary')
return auc_roc(y_true, y_score), auc_pr(y_true, y_score), f1
[docs]def ts_metrics(y_true, y_score):
"""
Calculates evaluation metrics for time series anomaly detection.
Args:
y_true (np.array, required):
Data label, 0 indicates normal timestamp, and 1 is anomaly.
y_score (np.array, required):
Predicted anomaly scores, higher score indicates higher likelihoods to be anomaly.
Returns:
tuple: A tuple containing:
- roc_auc_score (float):
The score of area under the ROC curve.
- average_precision_score (float):
The score of area under the precision-recall curve.
- best_f1 (float):
The best score of F1-score.
- best_p (float):
The best score of precision.
- best_r (float):
The best score of recall.
"""
best_f1, best_p, best_r = get_best_f1(y_true, y_score)
return auc_roc(y_true, y_score), auc_pr(y_true, y_score), best_f1, best_p, best_r
def get_best_f1(label, score):
"""
Return the best F1-score, precision and recall
Args:
label (np.array, required):
Data label, 0 indicates normal timestamp, and 1 is anomaly.
score (np.array, required):
Predicted anomaly scores, higher score indicates higher likelihoods to be anomaly.
Returns:
tuple: A tuple containing:
- best_f1 (float):
The best score of F1-score.
- best_p (float):
The best score of precision.
- best_r (float):
The best score of recall.
"""
precision, recall, _ = metrics.precision_recall_curve(y_true=label, probas_pred=score)
f1 = 2 * precision * recall / (precision + recall + 1e-5)
best_f1 = f1[np.argmax(f1)]
best_p = precision[np.argmax(f1)]
best_r = recall[np.argmax(f1)]
return best_f1, best_p, best_r
[docs]def ts_metrics_enhanced(y_true, y_score, y_test):
"""
This function calculates additional evaluation metrics for time series anomaly detection.
It returns a variety of metrics, including those sourced from the code in
[A Huet et al. KDD22] and [J Paparrizos et al. VLDB22].
The function requires three inputs: y_true (data label), y_score (predicted anomaly scores),
and y_test (predictions of events).
Args:
y_true (np.array):
Data label, where 0 indicates a normal timestamp and 1 indicates an anomaly.
y_score (np.array):
Predicted anomaly scores, where a higher score indicates a higher likelihood of being an anomaly.
y_test (np.array):
Predictions of events, where 0 indicates a normal timestamp and 1 indicates an anomaly.
Returns:
tuple: A tuple containing:
- auroc (float):
The score of the area under the ROC curve after point adjustment.
- aupr (float):
The score of the area under the precision-recall curve after point adjustment.
- best_f1 (float):
The best score of F1-score after point adjustment.
- best_p (float):
The best score of precision after point adjustment.
- best_r (float):
The best score of recall after point adjustment.
- affiliation_precision (float):
The score of affiliation precision.
- affiliation_recall (float):
The score of affiliation recall.
- vus_r_auroc (float):
The score of range VUS ROC.
- vus_r_aupr (float):
The score of range VUS PR.
- vus_roc (float):
The score of VUS ROC.
- vus_pr (float):
The score of VUS PR.
"""
auroc = auc_roc(y_true, point_adjustment(y_true, y_score))
aupr = auc_pr(y_true, point_adjustment(y_true, y_score))
best_f1, best_p, best_r = get_best_f1(y_true, point_adjustment(y_true, y_score))
events_pred = convert_vector_to_events(y_test)
events_gt = convert_vector_to_events(y_true)
Trange = (0, len(y_test))
affiliation = pr_from_events(events_pred, events_gt, Trange)
vus_results = get_range_vus_roc(y_score, y_true, slidingWindow=100) # default slidingWindow = 100
affiliation_precision = affiliation['Affiliation_Precision']
affiliation_recall = affiliation['Affiliation_Recall']
vus_r_auroc = vus_results["R_AUC_ROC"]
vus_r_aupr = vus_results["R_AUC_PR"]
vus_roc = vus_results["VUS_ROC"]
vus_pr = vus_results["VUS_PR"]
return auroc, aupr, best_f1, best_p, best_r, affiliation_precision, affiliation_recall, \
vus_r_auroc, vus_r_aupr, vus_roc, vus_pr