Examples¶

scikit-learn score polarity¶

from sklearn.ensemble import IsolationForest
from labelfree.metrics import relative_top_median_score

model = IsolationForest(random_state=0).fit(X)
scores = model.decision_function(X)

value = relative_top_median_score(
    scores,
    score_polarity="higher_is_normal",
)

Repeated-run stability¶

import numpy as np
from labelfree.metrics import top_k_stability_score

repeated_scores = np.vstack([scores_seed_1, scores_seed_2, scores_seed_3])
stability = top_k_stability_score(repeated_scores, top_fraction=0.05)

Rows must score the same samples in the same order.

IREOS candidate comparison¶

import numpy as np
from labelfree.metrics import ireos_scores

probabilities = np.vstack([lof_probabilities, knn_probabilities])
values = ireos_scores(X_scaled, probabilities)

Rows must contain detector-specific outlier probabilities in [0, 1]. ireos_scores selects one shared kernel-width range for a fair comparison.

Notebooks¶

• Metric tuning uses Optuna to swap label-free objectives across a shared detector pool and checks the selected configurations on a labeled holdout.
• Metric signal analysis compares label-free rankings with held-out detector quality across datasets and random splits.

Install notebook dependencies with:

python -m pip install "labelfree[examples]"

The notebooks contain no cached outputs, so displayed results always correspond to the installed package and dependencies.