This code used as a starting point an implementation of the paper "Efficient Search for Diverse Coherent Explanations" by Chriss Russel. Only the implementation of a mixed polytope input encoding is used here.
Original code by Chriss Russell can be found in the original repository.
He used Logistic Regression and focused on generating diverse explanations. This codebase used his formulation of mixed type inputs (polytopes) and modelling of the changes of inputs with respect to the original value of the factual. Other than that, code is my own. The code used was also significantly improved and clarified as to what the various parts of the code mean.
Instead of diverse coherent explanations, the focus here is on generating a set of counterfactuals closest to original factual.
Examples provided use the adult dataset (included), or MNIST for the mutli class problem.
The code uses the Gurobi solver for the MIP solver, and
gurobi-machinelearning
package for the NN computation.
There is also a custom NN implementation, using the methods presented by M. Fischetti and J. Jo
"Deep neural networks and mixed integer linear optimization"
That implementation has shown better performance when it comes to speed, but lower
quality of solutions, because it generates duplicate counterfactuals. If that does not
bother you, use the code in the custom_nn_implementation/ folder.
The encoder explicitly targets the FICO dataset and has made a couple of simple assumptions as to the form the dataset takes. Each variable is assumed to take a range of continuous values and a set of discrete values; as simplifying assumptions we assume that all strictly negative values are the discrete values, while the continuous values are the non-negative ones.
If you wish to add an entirely discrete variable i.e. without a continious range included, these variables should be indexed from zero. For example, in the adult dataset the 'workclass' variable takes the following values. {0: 'Government', -3: 'Other/Unknown', -2: 'Private', -1: 'Self-Employed'}
If this is not the case for your dataset, the code can be adapted to match assumptions, but it probably easier to manipulate the data so that it follows these assumptions -- this manipulation has already been done for the adult dataset.
The input encoder was improved from the work of Chriss Russel. The handling of categorical variables is corrected, so now the model works well for categorical, numerical and mixed input features.
This repository also contains a couple of attempts to create a utility function regarding the set of counterfactuals.
See example_objective.py for furhter details about the functions.
This is still a work-in-progress.
This repository is a part of Jiří Němeček's Master's Thesis at FEE CTU in Prague