Fairmodels

Fairmodels is an open-source R package that detects, visualises and helps mitigate bias in machine learning models. It works by taking one or more trained models that have already been wrapped as DALEX "explainer" objects (DALEX is a companion explainable-AI package from the same research group) and running them through a function called fairness_check(). 

The user supplies a protected attribute (such as gender or race) and designates which subgroup is "privileged." The package then computes model predictions for each subgroup and calculates twelve metrics derived from each subgroup's confusion matrix, including true positive rate, true negative rate, positive predictive value, false positive rate, statistical parity and F1 score. For each metric, it computes a "parity loss" value that measures how far each unprivileged subgroup's score diverges from the privileged subgroup's score; low parity loss on the relevant metrics indicates the model treats groups similarly (for example, low parity loss in both TPR and FPR together indicates "equalised odds" is satisfied). Results can be printed as a summary table or plotted to visually compare fairness across multiple models at once, which lets users benchmark different models or model versions against each other on the same fairness criteria.

Beyond detection, the package includes pre-processing techniques (adjusting training data before model fitting) and post-processing techniques (adjusting model outputs after fitting) that users can apply to reduce measured disparities. It also has a separate function, fairness_check_regression(), that approximates these same classification-style fairness metrics for regression models, since standard confusion-matrix metrics do not directly apply to continuous outputs. The package supports user-defined custom metrics as well.

Auto-discovered on 2026-02-11 by OECD Catalogue Automation