def _credit_clusters():
return distributions.Mixture(
components=[
_single_credit_cluster(idx) for idx in range(feature_dim)
],
weights=cluster_probs,
)
def two_group_credit_clusters(
group_likelihoods=(0.5, 0.5),
cluster_probabilities=DELAYED_IMPACT_CLUSTER_PROBS,
success_probabilities=DELAYED_IMPACT_SUCCESS_PROBS):
"""Returns a mixture of two credit cluster distributions.
Args:
group_likelihoods: Probabilities of choosing from each group.
Should be non-negative and sum to one.
cluster_probabilities: cluster_probabilities[i][j] gives the probability of
being drawn from credit cluster j conditioned on being from group i.
success_probabilities: success_probabilities[j] gives the probability of
successfully paying back a loan conditioned on being from being from
credit cluster j.
"""
components = []
for idx in [0, 1]:
# Use a one-hot encoding of group-id.
group_vec = np.zeros(2)
group_vec[idx] = 1
components.append(
_credit_cluster_builder(
group_membership=tuple(group_vec),
cluster_probs=tuple(cluster_probabilities[idx]),
success_probs=tuple(success_probabilities))())
return distributions.Mixture(components=components,
weights=group_likelihoods)
def test_mixture_returns_components(self):
my_distribution = distributions.Mixture(components=[
distributions.Constant((0, )),
distributions.Constant((1, ))
],
weights=[0.1, 0.9])
rng = np.random.RandomState(seed=100)
samples = [my_distribution.sample(rng) for _ in range(1000)]
self.assertSetEqual(set(samples), {(0, ), (1, )})
self.assertAlmostEqual(np.mean(samples), 0.9, delta=0.1)
def test_improper_distributions_raise_errors(self):
for p in [-10, -0.9, 1.3]:
with self.assertRaises(ValueError):
_ = distributions.Bernoulli(p=p)
for vec in [
[0.1, 0.3, 0.5], # Does not sum to one.
[0.5, 0.9, -0.4], # Has negative values.
]:
with self.assertRaises(ValueError):
_ = distributions.Mixture(
weights=vec,
components=[distributions.Constant(mean=(0, ))] * len(vec))
def _single_gmm():
"""Returns a mixture of gaussian applicant distributions."""
return distributions.Mixture(
components=[
ApplicantDistribution(
features=distributions.Gaussian(mean=mean, std=0.5),
group_membership=distributions.Constant(group),
will_default=distributions.Bernoulli(p=default_likelihoods[0])),
ApplicantDistribution(
features=distributions.Gaussian(
mean=np.array(mean) + np.array(intercluster_vec), std=0.5),
group_membership=distributions.Constant(group),
will_default=distributions.Bernoulli(p=default_likelihoods[1]))
],
weights=[0.3, 0.7])
def _rotated_gmm():
"""Returns a mixture of applicant distributions.
The second applicant distribution is a rotated version of the first.
"""
return distributions.Mixture(components=[
_gmm_applicant_builder(mean=mean,
intercluster_vec=intercluster_vec,
default_likelihoods=default_likelihoods,
group=[1, 0])(),
_gmm_applicant_builder(mean=_rotate(mean, degrees),
intercluster_vec=_rotate(
intercluster_vec, degrees),
default_likelihoods=default_likelihoods,
group=[0, 1])(),
],
weights=group_likelihoods)
def test_measure_distribution_change_measurement(self):
# The lower cluster has a 100% success rate and the upper cluster has a 0%
# success rate. This causes applicants to move constantly between clusters.
clusters = distributions.Mixture(
components=[
lending_params._credit_cluster_builder(
group_membership=[1, 0],
cluster_probs=[0.1, 0.9],
success_probs=[1.0, 0.0])(),
lending_params._credit_cluster_builder(
group_membership=[0, 1],
cluster_probs=[0.8, 0.2],
success_probs=[1.0, 0.0])(),
],
weights=(0.5, 0.5),
)
env = lending.DelayedImpactEnv(
lending_params.DelayedImpactParams(
applicant_distribution=clusters))
initial_distribution = lending_metrics.CreditDistribution(env, 0)
final_distribution = lending_metrics.CreditDistribution(env, -1)
# Giving a loan should change the distribution.
env.step(np.asarray(1))
# Take another step to move current state into history. This step does not
# change the distribution because the loan is rejected.
env.step(np.asarray(0))
self.assertEqual({
"0": [0.1, 0.9],
"1": [0.8, 0.2]
}, initial_distribution.measure(env))
self.assertNotEqual({
"0": [0.1, 0.9],
"1": [0.8, 0.2]
}, final_distribution.measure(env))
