Open sharathmalladi opened 5 years ago
Sharing more context on the current logic: When a counterfactual evaluation is created with a start and end date, we currently use the model file that was saved on the start date as input to vw to run the feature importance: https://github.com/microsoft/mwt-ds/blob/master/DataScience/ExperimentationAzure.py#L168
This is not the right thing if the model or policy was reset in between the evaluation period and the logic needs to be tweaked to adjust for that.
The feature importance should not use a model file that is prior to the policy reset date.