def test_sample_bootstrap_bandit_feedback():
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
bandit_feedback = dataset.obtain_batch_bandit_feedback()
bootstrap_bf = dataset.sample_bootstrap_bandit_feedback()
bf_keys = {"action", "position", "reward", "pscore", "context"}
for k in bf_keys:
assert len(bandit_feedback[k]) == len(bootstrap_bf[k])
bandit_feedback_timeseries: Dict = dataset.obtain_batch_bandit_feedback(
is_timeseries_split=True)[0]
bootstrap_bf_timeseries = dataset.sample_bootstrap_bandit_feedback(
is_timeseries_split=True)
for k in bf_keys:
assert len(bandit_feedback_timeseries[k]) == len(
bootstrap_bf_timeseries[k])
def test_sample_bootstrap_bandit_feedback():
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
bandit_feedback = dataset.obtain_batch_bandit_feedback()
bootstrap_bf = dataset.sample_bootstrap_bandit_feedback()
assert len(bandit_feedback["action"]) == len(bootstrap_bf["action"])
assert len(bandit_feedback["position"]) == len(bootstrap_bf["position"])
assert len(bandit_feedback["reward"]) == len(bootstrap_bf["reward"])
assert len(bandit_feedback["pscore"]) == len(bootstrap_bf["pscore"])
assert len(bandit_feedback["context"]) == len(bootstrap_bf["context"])
def test_sample_bootstrap_bandit_feedback():
with pytest.raises(ValueError):
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
dataset.sample_bootstrap_bandit_feedback(is_timeseries_split=True,
test_size=1.3)
with pytest.raises(ValueError):
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
dataset.sample_bootstrap_bandit_feedback(is_timeseries_split=True,
test_size=-0.5)
with pytest.raises(ValueError):
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
dataset.sample_bootstrap_bandit_feedback(sample_size=-50)
with pytest.raises(TypeError):
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
dataset.sample_bootstrap_bandit_feedback(sample_size=50.0)
with pytest.raises(ValueError):
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
dataset.sample_bootstrap_bandit_feedback(sample_size=10000000)
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
bandit_feedback = dataset.obtain_batch_bandit_feedback()
bootstrap_bf = dataset.sample_bootstrap_bandit_feedback()
bf_keys = {"action", "position", "reward", "pscore", "context"}
for k in bf_keys:
assert len(bandit_feedback[k]) == len(bootstrap_bf[k])
bandit_feedback_timeseries: Dict = dataset.obtain_batch_bandit_feedback(
is_timeseries_split=True)[0]
bootstrap_bf_timeseries = dataset.sample_bootstrap_bandit_feedback(
is_timeseries_split=True)
for k in bf_keys:
assert len(bandit_feedback_timeseries[k]) == len(
bootstrap_bf_timeseries[k])
sample_size = 1000
dataset = OpenBanditDataset(behavior_policy="random", campaign="all")
bootstrap_bf = dataset.sample_bootstrap_bandit_feedback(
sample_size=sample_size)
assert bootstrap_bf["n_rounds"] == sample_size
for k in bf_keys:
assert len(bootstrap_bf[k]) == sample_size
# compared OPE estimators
ope_estimators = [DirectMethod(), InverseProbabilityWeighting(), DoublyRobust()]
# a base ML model for regression model used in Direct Method and Doubly Robust
base_model = CalibratedClassifierCV(LogisticRegression(**hyperparams))
# ground-truth policy value of a counterfactual policy
# , which is estimated with factual (observed) rewards (on-policy estimation)
ground_truth_policy_value = OpenBanditDataset.calc_on_policy_policy_value_estimate(
behavior_policy=counterfactual_policy, campaign=campaign, data_path=data_path
)
evaluation_of_ope_results = {
est.estimator_name: np.zeros(n_boot_samples) for est in ope_estimators
}
for b in np.arange(n_boot_samples):
# sample bootstrap from batch logged bandit feedback
boot_bandit_feedback = obd.sample_bootstrap_bandit_feedback(random_state=b)
# run a counterfactual bandit algorithm on logged bandit feedback data
selected_actions = run_bandit_simulation(
bandit_feedback=boot_bandit_feedback, policy=policy
)
# evaluate the estimation performance of OPE estimators
ope = OffPolicyEvaluation(
bandit_feedback=boot_bandit_feedback,
action_context=obd.action_context,
regression_model=RegressionModel(base_model=base_model),
ope_estimators=ope_estimators,
)
relative_estimation_errors = ope.evaluate_performance_of_estimators(
selected_actions=selected_actions,
ground_truth_policy_value=ground_truth_policy_value,
)
)
start = time.time()
relative_ee = {
est.estimator_name: np.zeros(n_boot_samples)
for est in ope_estimators
}
for b in np.arange(n_boot_samples):
# load the pre-trained regression model
with open(reg_model_path / f"reg_model_{b}.pkl", "rb") as f:
reg_model = pickle.load(f)
with open(reg_model_path / f"is_for_reg_model_{b}.pkl", "rb") as f:
is_for_reg_model = pickle.load(f)
# sample bootstrap samples from batch logged bandit feedback
boot_bandit_feedback = obd.sample_bootstrap_bandit_feedback(
test_size=test_size,
is_timeseries_split=is_timeseries_split,
random_state=b)
for key_ in ["context", "action", "reward", "pscore", "position"]:
boot_bandit_feedback[key_] = boot_bandit_feedback[key_][
~is_for_reg_model]
if evaluation_policy == "bts":
policy = BernoulliTS(
n_actions=obd.n_actions,
len_list=obd.len_list,
is_zozotown_prior=
True, # replicate the policy in the ZOZOTOWN production
campaign=campaign,
random_state=random_state,
)
action_dist = policy.compute_batch_action_dist(
n_sim=100000, n_rounds=boot_bandit_feedback["n_rounds"])