new feature: is_factorizable option in SyntheticSlateBanditDataset

[aiueola]

new feature

• implemented is_factorizable option:
  • if is_factorizable=True, the action at each slot is sampled independently from other items (i.e., :math:pi(a_k | x)). Using this option, the actions in a slate may be duplicated. (newly implemented one to avoid computation cost to calculate pscore)
  • if is_factorizable=False, the action at each slot is sampled dependently on the former items (i.e., :math:pi(a_k | x, a_1, \ldots, a_{k-1})). Using this option, the actions in a slate will not be duplicated. (originally implemented one)

https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/obp/dataset/synthetic_slate.py#L97

• specifically, the following functions are changed to correspond is_factorizable option.

.sample_action_and_obtain_pscore()
https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/obp/dataset/synthetic_slate.py#L496

.obtain_pscore_given_evaluation_policy_logit()
https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/obp/dataset/synthetic_slate.py#L399

.calc_ground_truth_policy_value()
https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/obp/dataset/synthetic_slate.py#L815

.generate_evaluation_policy_pscore()
https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/obp/dataset/synthetic_slate.py#L884

refactor

• changed argname from evaluation_policy_logit to evaluation_policy_logit_ as there existed both evaluation_policy_logit and evaluation_policy_logit_ for the different functions with the same meaning.

tests

• add corresponding tests for is_factorizable option.

others

• minor fix on typos and docstrings.

[usaito]

@aiueola

Thanks!

• all impresessions -> all slates
  https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/tests/dataset/test_synthetic_slate.py#L370
  https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/tests/dataset/test_synthetic_slate.py#L417

• update remaining item... -> update the pscore given the remaining items...

• not factorizable -> nonfactorizable

• calculate marginal pscore -> calculate pscore_item_position

• factorizable_pscore -> pscore_when_factorizable

• How about getting if self.is_factorizable: out of the for loop as below? We don't have to check whether the policy is factorizable or not in the every loop, do we? Moreover, we don't have to define factorizable_pscore when self.is_factorizable=False.

# current
factorizable_pscore = softmax(evaluation_policy_logit_[i : i + 1])[0]            
for action_list in enumerated_slate_actions:                
    if self.is_factorizable:
        pscores.append(
            np.cumprod([factorizable_pscore[a_] for a_ in action_list])[-1]                   
        )
    else:

# my proposal
if self.is_factorizable:
    factorizable_pscore = softmax(evaluation_policy_logit_[i : i + 1])[0]        
    for action_list in enumerated_slate_actions:       
        pscores.append(
            np.cumprod([factorizable_pscore[a_] for a_ in action_list])[-1]                   
        )
else:

https://github.com/aiueola/zr-obp/blob/ea995f4337b94945baac7aa6547e36bd7e6d69d5/obp/dataset/synthetic_slate.py#L814-816