def _train_model(self):
"""Trains the internal model using all features seen so far."""
if not self._can_train():
return
self._threshold = np.clip(
threshold_policies.single_threshold(
predictions=self._features,
labels=self._labels,
weights=None,
cost_matrix=self._cost_matrix), _SCORE_MIN, _SCORE_MAX)
def test_weighted_single_threshold(self):
# With a vanilla cost matrix, this should result in accepting
# only predictions >= 0.75
predictions = np.array([0.25, 0.25, 0.33, 0.33, 0.75, 0.75])
labels = np.array([0, 1, 0, 1, 0, 1])
weights = np.array([3., 1., 2., 1., 1., 3.])
weights = weights / sum(weights)
vanilla_cost_matrix = params.CostMatrix(tn=0., fp=-1., fn=0., tp=1.)
weighted_threshold = threshold_policies.single_threshold(
predictions, labels, weights, vanilla_cost_matrix)
self.assertAlmostEqual(weighted_threshold, 0.75)
def test_weighted_single_threshold(self):
# Weighted labels/predictions are perfectly calibrated.
# With a vanilla cost matrix, this should result in threshold > 0.5
predictions = np.array([0.25, 0.25, 0.5, 0.5, 0.75, 0.75])
labels = np.array([0, 1, 0, 1, 0, 1])
weights = np.array([3., 1., 1., 1., 1., 3.])
weights = weights / sum(weights)
vanilla_cost_matrix = params.CostMatrix(tn=1., fp=-1., fn=-1., tp=1.)
weighted_threshold = threshold_policies.single_threshold(
predictions, labels, weights, vanilla_cost_matrix)
self.assertEqual(weighted_threshold, 0.75)
def _set_thresholds(self, training_corpus):
self.global_threshold = threshold_policies.single_threshold(
predictions=self._score_transform(training_corpus.get_features()),
labels=training_corpus.get_labels(),
weights=None,
cost_matrix=self.params.cost_matrix)
if self.params.threshold_policy == threshold_policies.ThresholdPolicy.EQUALIZE_OPPORTUNITY:
self.group_specific_thresholds = (
threshold_policies.equality_of_opportunity_thresholds(
group_predictions=self._recursively_apply_score_transform(
training_corpus.get_features(
stratify_by=self.params.group_key)),
group_labels=training_corpus.get_labels(
stratify_by=self.params.group_key),
group_weights=None,
cost_matrix=self.params.cost_matrix))