Reproduce “FairGBM: Gradient Boosting with Fairness Constraints” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
André F Cruz, Catarina Belém, Sérgio Jesus, João Bravo, Pedro Saleiro, Pedro Bizarro
FairGBM introduces an efficient in-processing framework for Gradient Boosted Decision Trees (GBDT) that incorporates fairness constraints. This approach enables models to achieve superior fairness-performance trade-offs and an order of magnitude faster training times compared to existing methods, making it suitable for high-stakes, tabular data applications.
Key insights
- FairGBM successfully adapts a dual ascent optimization framework to Gradient Boosted Decision Trees (GBDTs), providing an efficient in-processing method for fairness constraints that overcomes limitations of prior approaches.
- Using a proxy-Lagrangian formulation with smooth convex error rate proxies (based on cross-entropy loss) enables gradient-based optimization for non-differentiable fairness metrics, leading to a smoother training process.
- The empirical results demonstrate that the final deterministic model (last iterate) of FairGBM performs comparably to the theoretically guaranteed randomized classifier, simplifying practical deployment in real-world systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.