Controlled Model Debiasing through Minimal and Interpretable Updates

Authors: Federico Di Gennaro, Thibault Laugel, Vincent Grari, Marcin Detyniecki

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments, we demonstrate that our method achieves comparable fairness and accuracy performance to existing algorithmic fairness approaches, while requiring fewer prediction changes. Additionally, we show that COMMOD enables more meaningful and easier-to-understand prediction changes, enhancing its utility in practice. To summarize, our contributions are as follows: We validate its performance through experiments on classical fairness datasets, showcasing its debiasing efficacy and ability to perform fewer and more interpretable changes (Section 6).
Researcher Affiliation Collaboration Federico Di Gennaro1,2 EMAIL EPFL, Lausanne, Switzerland Thibault Laugel1 EMAIL AXA, Paris, France TRAIL, LIP6, Sorbonne Université, Paris, France Vincent Grari AXA, Paris, France TRAIL, LIP6, Sorbonne Université, Paris, France Marcin Detyniecki AXA, Paris, France TRAIL, LIP6, Sorbonne Université, Paris, France Polish Academy of Science, IBS PAN, Warsaw, Poland
Pseudocode Yes In this section we give a more detailed walk-through of our end-to-end training procedure for COMMOD, as well as a compact pseudocode listing. Algorithm 1 COMMOD Training (simplified pseudocode)
Open Source Code Yes Code to reproduce the experiments is available on the following repository: https://github.com/axa-rev-research/controlled-model-debiasing
Open Datasets Yes We experimentally validate two binary classification datasets, commonly used in the fairness literature (Hort et al., 2024): Law School (Wightman, 1998) and Compas (Angwin et al., 2016).
Dataset Splits Yes After splitting each dataset in Dtrain (70%) and Dtest (30%), we train our pretrained classifier f to optimize solely accuracy.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. It only mentions general computing environments without specific hardware.
Software Dependencies No In these experiments, we use a Logistic Regression classifier from the scikit-learn library, but any other classifier could be used since COMMOD and the proposed competitors are model-agnostic.
Experiment Setup Yes For COMMOD, we set a fixed value for the number of concepts k: 2 for Law School and 5 for Compas. Further details on implementation are available in Section C of the Appendix. ... The range of values we tested remained consistent across different datasets, with λfair 10, λratio 0.5, and λconcepts 1.