Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Efficient Fairness-Performance Pareto Front Computation

Authors: Mark Kozdoba, Binyamin Perets, Shie Mannor

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The approach was evaluated on several real world benchmark datasets and compares favorably to a number of recent state of the art fair representation and classification methods. We evaluate our approach on standard fairness benchmark datasets and compare its fairness-performance curve to multiple state-of-the-art fair representation methods.
Researcher Affiliation Collaboration 1Technion Israel Institute of Technology, Haifa, Israel 2NVIDIA
Pseudocode Yes The full algorithm is schematically show as Algorithm 1, Supplementary H.1.
Open Source Code Yes All evaluations can be found at https://github.com/bp6725/ Efficient-Fair-Pareto-Paper. The MIFPO algorithm s source code is available in the https://github.com/bp6725/Fair Pareto repository. The algorithm is also implemented as the "Fair Pareto" Python package on Py PI, which provides a scikit-learn compatible API for computing optimal fairness-performance Pareto fronts.
Open Datasets Yes Our experimental validation of MIFPO encompasses three standard fairness benchmarks: the Health dataset alongside two variants of ACSIncome one restricted to California (ACSIncome-CA) and another spanning the entire United States (ACSIncome-US). We conducted our evaluation using three of the most common datasets in this field, which are known to have relevance to real-world decision-making processes: the Adult dataset (income prediction), COMPAS (recidivism prediction), and LSAC (law school admission). The datasets used are publicly available.
Dataset Splits No For training the XGBoost model, a Grid Search CV approach is employed to find the best hyperparameters from a specified parameter grid, using 3-fold cross-validation. (The paper mentions 'test set' and '3-fold cross-validation' for hyperparameter tuning, but does not provide specific split percentages or methodology for the main experimental data splits for datasets like Health, ACSIncome, Adult, COMPAS, and LSAC.)
Hardware Specification Yes The experiments were conducted on a system with an Intel Core i9-12900KS CPU (16 cores, 24 threads), 64 GB of RAM, and an NVIDIA Ge Force RTX 3090 GPU.
Software Dependencies No For the calibrated classifier, we have used XGBoost (Chen et al., 2015), with Isotonic Regression calibration, as implemented in sklearn, Pedregosa et al. (2011). Next, as discussed in Sections 1, 4.1, MIFPO is a concave minimisation problem, under linear constraints. To solve its, we have used the DCCP framework and the associated solver (Shen et al., 2016, 2024), which are based on the combination of convexconcave programming (CCP) Lipp and Boyd (2016) and disciplined convex programming, Grant et al. (2006). (The paper mentions software like XGBoost, sklearn, and DCCP, but does not specify their version numbers.)
Experiment Setup Yes Throughout the experiments, we use the missclassification error loss h given by (4). As discussed in Section 4.1, we found it sufficient to use k = 5 throughout the paper. For training the XGBoost model, a Grid Search CV approach is employed to find the best hyperparameters from a specified parameter grid, using 3-fold cross-validation.