def calculate_bayes_opt(self, y0_prob, y1_prob, t_obs, t_obs_cf):
bayes_opt_pred_t0 = np.round(y0_prob)
bayes_opt_pred_t1 = np.round(y1_prob)
bayes_opt_pred_f = switch(bayes_opt_pred_t0, bayes_opt_pred_t1, t_obs)
bayes_opt_pred_cf = switch(bayes_opt_pred_t0, bayes_opt_pred_t1,
t_obs_cf)
bayes_opt_trmt_by_num = (y1_prob > y0_prob).astype(float)
bayes_opt_trmt_by_fcf = switch(y0_prob > y1_prob, y1_prob > y0_prob,
t_obs)
# bayes_opt_trmt_by_fcf = switch(1 - bayes_opt_trmt_by_num, bayes_opt_trmt_by_num, np.equal(bayes_opt_trmt_by_num > 0.5, t_obs > 0.5))
return bayes_opt_pred_f, bayes_opt_pred_cf, bayes_opt_trmt_by_num, bayes_opt_trmt_by_fcf
def choose_observed_treatment(self, x, a, z):
# A = 0 treatment
a0_t_obs_prob = np.expand_dims(
sigmoid(np.mean(x + z, axis=1) / self.outcome_scale), 1)
a0_t_obs = np.random.binomial(1, a0_t_obs_prob)
a0_t_obs_cf = 1 - a0_t_obs
# A = 1 treatment
a1_t_obs_prob = np.expand_dims(
sigmoid(np.mean(x - z, axis=1) / self.outcome_scale), 1)
a1_t_obs = np.random.binomial(1, a1_t_obs_prob)
a1_t_obs_cf = 1 - a1_t_obs
#combine outcomes by A
t_obs = switch(a0_t_obs, a1_t_obs, a)
t_obs_prob = switch(a0_t_obs_prob, a1_t_obs_prob, a)
t_obs_cf = switch(a0_t_obs_cf, a1_t_obs_cf, a)
return t_obs, t_obs_cf, t_obs_prob
def generate_outcomes(self, x, a, z):
mn = np.mean(x, axis=1)
logit_below_0 = mn / self.outcome_scale
logit_above_0 = (mn + self.offset) / self.outcome_scale
logits = switch(logit_below_0, logit_above_0, mn > 0)
y0_prob = np.expand_dims(sigmoid(logits), 1)
y1_prob = np.expand_dims(sigmoid(-logits), 1)
y0 = np.random.binomial(1, y0_prob)
y1 = np.random.binomial(1, y1_prob)
return y0, y1, y0_prob, y1_prob
def generate_factual_and_counterfactual_data(self):
#everything the same, but then select observed A at the end
#make dataset
x0, a0, z0 = self.generate_factual_data(self.mu0, self.sd0,
self.num_data, self.xdim, 0)
x0_cf, a0_cf, z0_cf = self.generate_counterfactual_data(
x0, a0, z0, self.mu0, self.sd0, self.mu1, self.sd1)
dat0 = self.get_treatment_outcome_data(x0, a0, z0)
dat0_cf = self.get_treatment_outcome_data(x0_cf, a0_cf, z0_cf)
#select observed sensitive attribute as a function of Z
observed_a = self.choose_observed_sensitive_attribute(z0)
#save data
save_dict_f = {
k: switch(dat0[k], dat0_cf[k], observed_a)
for k in dat0
}
save_dict_cf = {
k: switch(dat0[k], dat0_cf[k], 1 - observed_a)
for k in dat0
}
return save_dict_f, save_dict_cf
def generate_outcomes(self, x, a, z):
#A = 0 outcomes
a0_y0_prob = np.expand_dims(
sigmoid(np.mean(x + z, axis=1) / self.outcome_scale), 1)
a0_y1_prob = np.expand_dims(
sigmoid(-np.mean(x - z, axis=1) / self.outcome_scale), 1)
a0_y0 = np.random.binomial(1, a0_y0_prob)
a0_y1 = np.random.binomial(1, a0_y1_prob)
#A = 1 outcomes
a1_y0_prob = np.expand_dims(
sigmoid(np.mean(x - z, axis=1) / self.outcome_scale), 1)
a1_y1_prob = np.expand_dims(
sigmoid(-np.mean(x + z, axis=1) / self.outcome_scale), 1)
a1_y0 = np.random.binomial(1, a1_y0_prob)
a1_y1 = np.random.binomial(1, a1_y1_prob)
#combine outcomes by A
y0 = switch(a0_y0, a1_y0, a)
y1 = switch(a0_y1, a1_y1, a)
y0_prob = switch(a0_y0_prob, a1_y0_prob, a)
y1_prob = switch(a0_y1_prob, a1_y1_prob, a)
return y0, y1, y0_prob, y1_prob
def calculate_bayes_opt(y0_prob, y1_prob, t_obs, t_obs_cf):
bayes_opt_pred_t0 = np.round(y0_prob)
bayes_opt_pred_t1 = np.round(y1_prob)
bayes_opt_pred_f = switch(bayes_opt_pred_t0, bayes_opt_pred_t1, t_obs)
bayes_opt_pred_cf = switch(bayes_opt_pred_t0, bayes_opt_pred_t1, t_obs_cf)
return bayes_opt_pred_f, bayes_opt_pred_cf