kenchi/metrics.py
import numpy as np
from sklearn.metrics import auc, recall_score
from sklearn.utils import check_random_state
__all__ = ['LeeLiuScorer', 'NegativeMVAUCScorer']
def _lebesgue_measure(score_samples, offset, data_volume):
"""Compute Lebesgue measure."""
return np.mean(score_samples >= offset) * data_volume
class LeeLiuScorer:
"""Lee-Liu scorer.
References
----------
.. [#lee03] Lee, W. S, and Liu, B.,
"Learning with positive and unlabeled examples using weighted Logistic
Regression,"
In Proceedings of ICML, pp. 448-455, 2003.
"""
def __call__(self, det, X, y=None):
"""Compute the Lee-Liu metric.
Parameters
----------
det : object
Detector.
X : array-like of shape (n_samples, n_features), default None
Data.
y : array-like of shape (n_samples,), default None
Labels. If None, assume that all samples are positive.
Returns
-------
score : float
Lee-Liu metric.
"""
y_pred = det.predict(X)
if y is None:
y = np.ones_like(y_pred)
r = recall_score(y, y_pred)
return r ** 2 / (1. - det.contamination_)
class NegativeMVAUCScorer:
"""Negative MV AUC scorer.
Parameters
----------
data_max : array-like of shape (n_features,)
Per feature maximum seen in the data.
data_min : array-like of shape (n_features,)
Per feature minimum seen in the data.
interval : tuple, default (0.9, 0.999)
Interval of probabilities.
n_offsets : int, default 1000
Number of offsets.
n_uniform_samples : int, default 1000
Number of samples which are drawn from the uniform distribution over
the hypercube enclosing the data.
random_state : int or RandomState instance, default None
Seed of the pseudo random number generator.
References
----------
.. [#goix16] Goix, N.,
"How to evaluate the quality of unsupervised anomaly detection
algorithms?"
In ICML Anomaly Detection Workshop, 2016.
"""
def __init__(
self, data_max, data_min, interval=(0.9, 0.999),
n_offsets=1000, n_uniform_samples=1000, random_state=None
):
self.data_max = data_max
self.data_min = data_min
self.interval = interval
self.n_offsets = n_offsets
self.n_uniform_samples = n_uniform_samples
self.random_state = random_state
self.internal_state = check_random_state(random_state).get_state()
def __call__(self, det, X, y=None):
"""Compute the opposite of the area under the Mass-Volume (MV) curve.
Parameters
----------
det : object
Detector.
X : array-like of shape (n_samples, n_features)
Data.
y : ignored
Returns
-------
score : float
Opposite of the area under the MV curve.
"""
rnd = np.random.RandomState()
rnd.set_state(self.internal_state)
U = rnd.uniform(
low = self.data_min,
high = self.data_max,
size = (self.n_uniform_samples, det.n_features_)
)
score_samples = det.score_samples(X)
score_uniform_samples = det.score_samples(U)
mass, volume, _ = self._mv_curve(
score_samples, score_uniform_samples
)
is_in_range = \
(self.interval[0] <= mass) & (mass <= self.interval[1])
return -auc(mass[is_in_range], volume[is_in_range], reorder=True)
def _mv_curve(self, score_samples, score_uniform_samples):
"""Compute mass-volume pairs for different offsets.
Parameters
----------
score_samples : array-like of shape (n_samples,)
Opposite of the anomaly score for each sample.
score_uniform_samples : array-like of shape (n_uniform_samples,)
Opposite of the anomaly score for each sample which is drawn from
the uniform distribution over the hypercube enclosing the data.
Returns
-------
mass : array-like of shape (n_offsets,)
volume : array-like of shape (n_offsets,)
offsets : array-like of shape (n_offsets,)
"""
data_volume = np.prod(self.data_max - self.data_min)
mass = np.linspace(0., 1., self.n_offsets)
offsets = np.percentile(
score_samples, 100. * (1. - mass)
)
volume = np.vectorize(
_lebesgue_measure, excluded=[0, 2]
)(score_uniform_samples, offsets, data_volume)
return mass, volume, offsets