supervised-discretization

This repository contains the code for the paper Supervised Feature Compression based on Counterfactual Analysis

Installation

• The MILP problem for computing the Counterfactual Explanation for a point is implemented in Gurobi. An active Gurobi Licence is needed to run the code.

• The package can be installed with the command:

pip install SupervisedDiscretization

Hyperparameters

The implementation of the FCCA procedure can be found in the file discretize.py that contains the Python class FCCA which takes the following parameters:

• estimator: an unfitted binary classifier from the sklearn package. It can be one of the following: RandomForestClassifier, GradientBoosting, LinearSVC, SVC(kernel='linear'). It is also possible to take in input GridSearchCV to choose in cross validation the parameters of the estimator;
• p0, p1: lower and upper bound for the classification probability of points for which computing the Counterfactual Explanation;
• lambda0, lambda1, lambda2: hyperparameters for the Counterfactual Explanation problem that represents respectively the weights for the l0-, l1- and l2- norm;
• compress: boolean that is set to True to merge thresholds whose absolute difference is smaller than 0.01;
• timelimit: time limit in seconds for solving the Counterfactual Explanations problem;
• verbose: boolean that is set to True to print some informations about the process of fitting the FCCA procedure.

The FCCA class offers the following methods:

• fit: method for fitting the FCCA procedure;
• transform: method for discretizing a dataset by using the set of thresholds previously computed via the fit method;
• fit_transform: method for applying in sequence the fit and transform methods;
• selectThresholds: method for setting a different value of Q after the fit has been called; this method allows to subsample the set of thresholds in a fast way without recomputing the FCCA procedure.

Execution

We report an example on how to use the FCCA procedure on new data. The example can also be found in the file example.py

import pandas as pd
from sklearn.ensemble import GradientBoostingClassifier
from SupervisedDiscretization.discretizer import FCCA

if __name__ == '__main__':
    # Reading the dataset
    data = pd.read_csv('datasets/boston.csv')
    label_column = data.columns[-1]
    feature_columns = data.columns[:-1]

    # Train - test split
    data_ts = data.sample(n=int(0.3*len(data)))
    data_tr = data.drop(index=data_ts.index)

    x_tr, y_tr = data_tr[feature_columns], data_tr[label_column]
    x_ts, y_ts = data_ts[feature_columns], data_ts[label_column]

    # Target model
    target = GradientBoostingClassifier(max_depth=2, n_estimators=100,learning_rate=0.1)

    # Hyperparameters for the discretization - default values
    discretizer = FCCA(target, p0=0.5, p1=1, lambda0=0.1, lambda1=1, lambda2=0)

    # Discretization
    x_tr_discr, y_tr_discr = discretizer.fit_transform(x_tr, y_tr)
    x_ts_discr, y_ts_discr = discretizer.transform(x_ts, y_ts)

    # Compression - inconsistency rate
    print(f'Compression rate: {discretizer.compression_rate(x_ts, y_ts)}')
    print(f'Inconsistency rate: {discretizer.inconsistency_rate(x_ts, y_ts)}')

    print('Setting Q to 0.7')
    # Increasing the value of Q
    tao_q = discretizer.selectThresholds(0.7)

    # Discretization
    x_tr_discr, y_tr_discr = discretizer.transform(x_tr, y_tr, tao_q)
    x_ts_discr, y_ts_discr = discretizer.transform(x_ts, y_ts, tao_q)

    # Compression - inconsistency rate
    print(f'Compression rate: {discretizer.compression_rate(x_ts, y_ts, tao_q)}')
    print(f'Inconsistency rate: {discretizer.inconsistency_rate(x_ts, y_ts, tao_q)}')