Evaluation

Install evaluation extras first:

pip install "aberrant[eval]"

Common metrics

• average_precision_score (recommended for imbalanced anomaly data)
• roc_auc_score (useful secondary metric)

Example

from sklearn.metrics import average_precision_score, roc_auc_score

labels, scores = [], []
for x, y in dataset.stream():
    model.learn_one(x)
    labels.append(y)
    scores.append(model.score_one(x))

print("PR-AUC:", average_precision_score(labels, scores))
print("ROC-AUC:", roc_auc_score(labels, scores))

Streaming pitfalls

• Separate warmup from evaluation.
• Do not leak future labels into threshold calibration.
• Compare models under the same stream order and seed.