Select subset of features for perturbations [Counterfactuals Guided by Prototypes]

[gibranhasan14]

Hello,

I'm new to alibi and am loving it so far! I've been going through the Counterfactual Proto examples and wanted to ask how I could specify which input features should be perturbed for the counterfactual instances?

My dataset includes many features (Categorial & Numerical), but only a select few can realistically be altered.

Using the Counterfactual explanations with one-hot encoded categorical variables for the adult dataset as an example:

After training and fitting the Neural Net, if I only wanted to select features "Education" and "Age" as the instances to attack, what changes would I need to make? Would it make sense to just select the desired features in the sample instance from X_train?

[jklaise]

Hi, this is something we need to document more clearly but there are also some bugs preventing the proper use of this feature. -

Freezing certain features should be done via setting the feature_range parameter to be a tuple of 2 arrays specifying the minimum and maximum allowed values feature-wise. By setting the minimum and maximum to be equal to the feature value you can effectively "freeze" the feature.

However, there is a subtle bug in the current implementation that doesn't actually take into account the feature range (thanks for reminding me about this!) which we need to fix. Also, the API is quite clunky because for categorical features you would have to set the min and max to the value of the numerical encoding which you don't know a priori as this is calculated inside the method. This needs to change in the future.