In LimeTabularExplainer(), if we set sample_around_instance = True, then it sets data = data + instance. (in __data_inverse() call, after explain_instance() call).
But then when it returns data, inverse, we do
scaleddata = (data - self.scalar.mean)/self.scalar.scale_.
But self.scalar.mean_ stores the mean of the feature over the training sample (self.scaler.fit(training_data)) and not the instance, so are we re-standardizing the data correctly in this case?
shouldn't we standardize with the instance here? or at least with the sample mean of the generated sample?
In LimeTabularExplainer(), if we set sample_around_instance = True, then it sets data = data + instance. (in __data_inverse() call, after explain_instance() call).
But then when it returns data, inverse, we do scaleddata = (data - self.scalar.mean)/self.scalar.scale_.
But self.scalar.mean_ stores the mean of the feature over the training sample (self.scaler.fit(training_data)) and not the instance, so are we re-standardizing the data correctly in this case?
shouldn't we standardize with the instance here? or at least with the sample mean of the generated sample?