Fair-Seldonian

Fairness-constrained machine learning with high-confidence guarantees.

Fair-Seldonian is a Python framework implementing the Quasi-Seldonian Algorithm (QSA) for training ML models that provably satisfy fairness constraints. Given a behavioral constraint and a confidence level \(\delta\), the algorithm either returns a model satisfying the constraint with probability \(\geq 1 - \delta\), or returns No Solution Found — never an unsafe model.

Documentation	parulgupta1004.github.io/fair-seldonian
Repository	github.com/parulgupta1004/fair-seldonian
Paper	Thomas et al., Science 366 (2019) — doi:10.1126/science.aag3311
Python	3.10+

Note

For the foundational work on Seldonian algorithms, see:

Thomas, P.S., da Silva, B.C., Barto, A.G., Giguere, S., Brun, Y., & Brunskill, E. (2019). “Preventing undesirable behavior of intelligent machines.” Science, 366(6468), 999–1004. [DOI] [Project site]

User Guide

• Introduction
  • Motivation
  • Problem Formulation
  • Framework Overview
  • Constraint Specification
  • Confidence Bound Propagation
• Getting Started
  • Installation
  • Running Experiments
  • Library Usage
  • Configuration
  • Extending the Framework

Algorithm Details

• Mathematical Background
  • Notation
  • Quasi-Seldonian Algorithm
  • Delta Splitting
  • Interval Arithmetic
  • Predicted Bounds
• Algorithm Variants
  • Baseline QSA (base)
  • Modified Confidence Interval (mod)
  • Constant-Aware Delta Allocation (const)
  • Union Bound Optimization (bound)
  • All Optimizations (opt)
  • Lagrangian/KKT Optimization

API Reference

• fair_seldonian package
  • Subpackages
    • fair_seldonian.algorithms package
    • fair_seldonian.constraints package
    • fair_seldonian.data package
    • fair_seldonian.experiments package
    • fair_seldonian.models package
  • Submodules
  • fair_seldonian.config module
    • SeldonianConfig
  • Module contents

Bibliography

• Bibliography

Indices and Tables

• Index

• Module Index