[Paper Review] Causal Intersectionality and Fair Ranking

Ke Yang, Joshua R. Loftus+1 more|arXiv (Cornell University)|Jun 15, 2020

Game Theory and Voting Systems43 references17 citations

TL;DR

This paper proposes a causal modeling framework for intersectional fairness in ranking systems, using counterfactual scores to mitigate bias across overlapping sensitive attributes like race and gender. By computing what an individual's score would be if they belonged to a different intersectional group, the method achieves fairer top-k rankings, as demonstrated on real and synthetic datasets with improved representation for underrepresented subgroups.

ABSTRACT

In this paper we propose a causal modeling approach to intersectional fairness, and a flexible, task-specific method for computing intersectionally fair rankings. Rankings are used in many contexts, ranging from Web search to college admissions, but causal inference for fair rankings has received limited attention. Additionally, the growing literature on causal fairness has directed little attention to intersectionality. By bringing these issues together in a formal causal framework we make the application of intersectionality in algorithmic fairness explicit, connected to important real world effects and domain knowledge, and transparent about technical limitations. We experimentally evaluate our approach on real and synthetic datasets, exploring its behavior under different structural assumptions.

Motivation & Objective

• To address the gap in causal fairness research for ranking tasks, particularly in the context of intersectional bias.
• To formalize intersectional fairness using structural causal models (SCMs) and counterfactual reasoning.
• To develop a flexible, task-specific method for computing fair rankings that accounts for multiple sensitive attributes and their intersections.
• To evaluate the method empirically on real and synthetic datasets under varying structural assumptions.
• To connect intersectional fairness to causal fairness literature, grounding fairness in real-world causal relationships and domain knowledge.

Proposed method

• The method uses a structural causal model (SCM) with a directed acyclic graph (DAG) to represent causal relationships among variables, including sensitive attributes (e.g., gender, race), mediators, and outcome scores.
• Counterfactual scores are computed by do-calculus interventions: for each individual, the model simulates how their score would change if they belonged to a different intersectional subgroup (e.g., Black woman instead of White man).
• Rankings are then generated based on these counterfactual scores, ensuring fairness across intersectional groups by removing the causal effect of sensitive attributes.
• The approach distinguishes between resolving and non-resolving mediators, allowing nuanced modeling of which intermediate variables are ethically justifiable for decision-making.
• The method supports both score-based ranking and learning-to-rank (LTR) tasks, with experiments using ListNet for LTR and counterfactual training.
• The framework is implemented in code and publicly available at https://github.com/DataResponsibly/CIFRank.

Experimental results

Research questions

• RQ1 How can intersectional fairness be formalized within a causal framework for ranking systems?
• RQ2 To what extent can counterfactual reasoning reduce bias across overlapping sensitive attributes like race and gender in top-k rankings?
• RQ3 How does the method perform under different structural assumptions, such as the presence of confounders or mediators?
• RQ4 Can the method achieve demographic parity while minimizing utility loss across intersectional groups?
• RQ5 How does the choice of resolving vs. non-resolving mediators affect fairness outcomes?

Key findings

• The counterfactual ranking method successfully improved representation of small and geographically underrepresented departments in the CSRankings dataset, moving small departments into the top-20 and achieving greater geographic balance.
• In the moving company and COMPAS datasets, the method reduced demographic disparity in selection rates across intersectional groups (e.g., Black women, White men), with the resolving cf-LTR variant showing the most balanced outcomes.
• The method achieved reasonable fairness improvements even when utility loss was imbalanced across intersectional groups, with observed imbalances ranging from 1% to 5% in experiments.
• The approach demonstrated robustness in LTR tasks using ListNet, with the learned ranker showing insensitivity to counterfactual transformations, suggesting potential for test-time fairness guarantees.
• The method’s performance was sensitive to assumptions about sequential ignorability, particularly in models with multiple mediators or unobserved confounders, highlighting the importance of model validity.
• The framework enables transparent, interpretable fairness interventions grounded in causal structure, supporting inclusive decision-making processes.