def calc_covariance(s, decisions, ips_weights):
new_s = (2 * s) - 1
if ips_weights is not None:
mu_s = mean(new_s * ips_weights, axis=0)
d = decisions * ips_weights
else:
mu_s = mean(new_s, axis=0)
d = decisions
covariance = (new_s - mu_s) * d
return covariance
def fairness_function(type=None, **fairness_kwargs):
s = fairness_kwargs["s"]
ips_weights = fairness_kwargs[
"ips_weights"] if "ips_weights" in fairness_kwargs else None
decisions = "decision_probabilities" not in fairness_kwargs or not args.policy_type == "NN"
decisions = fairness_kwargs["decisions"] if decisions else fairness_kwargs[
"decision_probabilities"]
y = fairness_kwargs["y"]
type = args.fairness_type if type is None else type
if type == "BD_DP":
benefit = calc_benefit(decisions, ips_weights)
return mean_difference(benefit, s)
elif type == "COV_DP":
covariance = calc_covariance(s, decisions, ips_weights)
return mean(covariance, axis=0)
elif type == "BD_EOP":
benefit = calc_benefit(decisions, ips_weights)
y1_indices = np.where(y.squeeze() == 1)
return mean_difference(benefit[y1_indices], s[y1_indices])
def fairness_function_gradient(**fairness_kwargs):
policy = fairness_kwargs["policy"]
x = fairness_kwargs["x"]
s = fairness_kwargs["s"]
y = fairness_kwargs["y"]
decisions = fairness_kwargs["decisions"]
ips_weights = fairness_kwargs["ips_weights"]
if args.fairness_type == "BD_DP" or args.fairness_type == "BD_EOP":
result = calc_benefit(decisions, ips_weights)
elif args.fairness_type == "COV_DP":
result = calc_covariance(s, decisions, ips_weights)
log_gradient = policy.log_policy_gradient(x, s)
grad = log_gradient * result
if args.fairness_type == "BD_DP":
return mean_difference(grad, s)
elif args.fairness_type == "COV_DP":
return mean(grad, axis=0)
elif args.fairness_type == "BD_EOP":
y1_indices = np.where(y == 1)
return mean_difference(grad[y1_indices], s[y1_indices])
def mean(self):
return mean(self._measure, axis=1)
def cost_utility_probability(cost_factor=0.5, **utility_parameters):
utility = _utility_core(cost_factor, True, **utility_parameters)
return mean(utility, axis=0)
def cost_utility_gradient(cost_factor=0.5, **utility_parameters):
utility = _utility_core(cost_factor, False, **utility_parameters)
log_policy_gradient = utility_parameters["policy"].log_policy_gradient(
utility_parameters["x"], utility_parameters["s"])
utility_grad = log_policy_gradient * utility
return mean(utility_grad, axis=0)