mmm-fair
Multi-attribute, Multi-objective, Multi-definition aware fair classifiers.
mmm-fair is a package of boosting based multi-fair classifiers that promotes learning fairness-aware predictions under class-imbalance w.r.t. multiple protected attribute and capable to work with varities of fairness definition.
🎏 Multi-attribute
🛠️ Multi-objective: Uses post processing MOO selection to deliver an ensemble that deliver fair, accurate, and class balanced predictions
⚙️ MAMMOth Workflow integration (optionally can transform model into onnx format, making it plugable to MAMMOth)
📈 Produces various pareto plots to give insight to user about what are different ensemble points are capable of delivering and
🧩 User can accordingly update the model based on solution weight displayed in the pareto plot and set the model that fits their criteria the best.
mmm-fair is a library of fairness-aware classifiers that when trained on a given dataset, a fairness objective (e.g., demographic parity), and a set of protected attributes, provides a fair classifier learning a multi-objective trade-off among accuracy, class-imbalance, and multi-fairness (fairness w.r.t. multiple protected attributes). Further, provides a Pareto exploration showing the trade-off points available between the: 1) multiple training objectives, 2) fairness for multiple protected attributes, and 3) multiple fairness definitions.