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..

FairICP: Encouraging Equalized Odds via Inverse Conditional Permutation

Authors: Yuheng Lai, Leying Guan

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The efficacy and adaptability of our method are demonstrated through both simulation studies and empirical analyses of real-world datasets.
Researcher Affiliation Academia 1Department of Statistics, University of Wisconsin-Madison 2Department of Biostatistics, Yale University. Correspondence to: Leying Guan <EMAIL>.
Pseudocode Yes Algorithm 1 Fair ICP: Fairness-aware learning via inverse conditional permutation Input: Data (X, A, Y) = {(Xi, Ai, Yi)}i Itr Parameters: penalty weight µ, step size α, number of gradient steps Ng, and iterations T. Output: predictive model ˆfˆθf ( ) and discriminator ˆDˆθd( ).
Open Source Code No The paper discusses implementing Fair ICP and other methods. It mentions adapting code for FDL from 'https://github.com/yromano/fair_dummies', HGR from 'https://github.com/criteo-research/continuous-fairness', Gerry Fair from 'https://github.com/algowatchpenn/Gerry Fair', and Reduction from 'https://github.com/fairlearn/fairlearn'. However, there is no explicit statement or link provided for the open-source code of Fair ICP itself.
Open Datasets Yes Communities and Crime Data Set: This dataset contains 1994 samples and 122 features. The goal is to build a regression model predicting the continous violent crime rate. ACSIncome Dataset: We use the ACSIncome dataset from Ding et al. (2021) with 100,000 instances (subsampled) and 10 features. Adult Dataset: The dataset consists of 48,842 instances and the task is the same as ACSIncome. COMPAS Dataset: The Pro Publica s COMPAS recidivism dataset contains 5278 examples and 11 features (Fabris et al., 2022).
Dataset Splits Yes For all experiments, the data is repeated divided into a training set (60%) and a test set (40%) 100 times, with the average results on the test sets reported . We set K {1, 5, 10}, ω {0.6, 0.9} to investigate different levels of dependence on A, and the sample size for training/test data to be 500/400.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory details) used for running its experiments.
Software Dependencies Yes For KPC (Huang et al., 2022), we use R Package KPC (Huang, 2022) with the default Gaussian kernel and other parameters. Huang, Z. KPC: Kernel Partial Correlation Coefficient, 2022. URL https://CRAN.R-project.org/ package=KPC. R package version 0.1.2.
Experiment Setup Yes For all the models evaluated (Fair ICP, Fair CP, FDL, Oracle), we set the hyperparameters as follows: We set f as a linear model and use the Adam optimizer with a mini-batch size in {16, 32, 64}, learning rate in {1e-4, 1e-3, 1e-2}, and the number of epochs in {20, 40, 60, 80, 100, 120, 140, 160, 180, 200}. The discriminator is implemented as a four-layer neural network with a hidden layer of size 64 and Re LU non-linearities. We use the Adam optimizer, with a fixed learning rate of 1e-4. For the MAF used to estimate the conditional density (Y |A and A|Y ) in the training phase, we use MAF with one MADE component and one hidden layer with 2 conditional inputs nodes, and for optimizer we choose Adam with 0.01 l1 penalty and 0.1 learning rate.