GeCo: Quality Counterfactual Explanations in Real Time

To run GeCo, you need to have Julia installed (link). Then you can run the following commands to load the package.

using Pkg; Pkg.activate(".")
using GeCo

We provide scripts to load the data and model. For instance:

include("scripts/credit/credit_setup_MACE.jl");
include("scripts/adult/adult_setup_MACE.jl");
include("scripts/yelp/yelp_setup_PRF.jl");
include("scripts/allstate/allstate_setup_PRF.jl");

Then run the following command to compute the explanations:

explanation,  = @time explain(orig_entity, X, path, classifier)

where orig_entity is the entity to be explained, X is the dataset, path is the path to the directory to the dataset, classifier is the model that is explained.

The explain function accepts the following optional keyword parameters:

• desired_class -- value of the desired output (default = 1)
• k (Int64) -- number of selected candidates during selection (default = 100)
• max_num_generations (Int64) -- maximum number of generations for genetic algorithm (default = 100)
• min_num_generations (Int64) -- minimum number of generations for genetic algorithm (default = 3)
• max_num_samples (Int64) -- maximum number of samples during mutation (default = 5)
• convergence_k (Int64) -- k used to check convergence (default = 5)
• norm_ratio -- parameters used for the distance function ([0.25, 0.25, 0.25, 0.25])
• domains -- precomputed active domain for each feature group (default = nothing)
• compress_data (Bool) -- use Delta representation (default = true)
• return_df (Bool) -- return a DataFrame instead of compressed Delta representation (default = false)
• verbose (Bool) -- verbose output (default = false)

To print out the proposed actions for the top counterfactuals run:

actions(explanation, orig_instance)

If you would like to precompute the domains run:

domains = initDomains(path, X)