A Generic Framework for Conformal Fairness
Aditya T. Vadlamani, Anutam Srinivasan, Pranav Maneriker, Ali Payani, Srinivasan Parthasarathy
TL;DR
This paper introduces Conformal Fairness, a framework that extends conformal prediction to ensure fairness across sensitive groups, applicable to non-IID data like graphs, with theoretical guarantees and empirical validation.
Contribution
It formalizes Conformal Fairness, providing a new algorithm that controls coverage gaps between sensitive groups under exchangeability, broadening conformal prediction's applicability.
Findings
The algorithm effectively controls fairness gaps in experiments.
The framework applies to non-IID data such as graphs.
Theoretical guarantees align with empirical results.
Abstract
Conformal Prediction (CP) is a popular method for uncertainty quantification with machine learning models. While conformal prediction provides probabilistic guarantees regarding the coverage of the true label, these guarantees are agnostic to the presence of sensitive attributes within the dataset. In this work, we formalize \textit{Conformal Fairness}, a notion of fairness using conformal predictors, and provide a theoretically well-founded algorithm and associated framework to control for the gaps in coverage between different sensitive groups. Our framework leverages the exchangeability assumption (implicit to CP) rather than the typical IID assumption, allowing us to apply the notion of Conformal Fairness to data types and tasks that are not IID, such as graph data. Experiments were conducted on graph and tabular datasets to demonstrate that the algorithm can control…
Peer Reviews
Decision·ICLR 2025 Poster
The paper's main asset is a reasonably broad evaluation of the proposed methods on a variety of datasets where fairness concerns may necessitate using one of the tabulated conformal fairness metrics. Compared to prior works the experimental section appears more extensive, both in terms of metrics and datasets. Moreover, conformal fairness guarantees were displayed on graph data in addition to the standard supervised tasks. The results indicate that explicitly enforcing fairness according to any
--- The main weakness and bottleneck at this point is the writing of the manuscript. In particular, the writing of Section 3, which introduces the objectives, the general algorithm, and the analysis, is currently not acceptable. Indeed, lots of key notions and terms, both on the fairness side and on the conformal side, are not properly and unambiguously introduced and/or are discussed in arbitrary order. (Fairness: groups/group collections are never formally defined, nor are requirements on the
- Originality: Introduces "Conformal Fairness," extending Conformal Prediction (CP) to address fairness in uncertainty quantification, especially for non-IID data. - Quality: Provides strong theoretical backing with rigorous proofs and effective validation through experiments on graph and tabular datasets. - Clarity: Clearly defined theoretical concepts and a stepwise algorithm description make the methodology accessible to those with relevant background knowledge. Providing additional backgroun
- The paper lacks a detailed discussion of the fairness-efficiency trade-off. Quantifying acceptable efficiency losses when fairness is improved would make the results more actionable for practitioners balancing both aspects. - The extension of the exchangeability assumption to real-world data may not always hold. Adding empirical evidence or discussion on when this assumption is valid in practice would make the claims more robust. -Lack of comparison with all existing fairness-aware methods l
Conformal prediction is a very interesting area and there are many interesting works in this area. It is a natural question to ask how to achieve fairness for models in this setting. The authors provide a very general framework that can achieve many different definitions of algorithmic fairness. The paper also did many interesting experiments in graph datasets.
I personally think the author does not elaborate enough on why they chose this algorithm design. The writing of this paper could be improved.
Code & Models
Videos
Taxonomy
TopicsEthics and Social Impacts of AI · Advanced Graph Neural Networks · Adversarial Robustness in Machine Learning