Intra Group Rank Preservation in Fair Machine Learning

108 concepts47 papers1 proposal

Imagine you are designing an algorithmic hiring system. You want to ensure the system is fair, so you apply a standard group fairness intervention like demographic parity to equalize the selection rates between a majority group and an underrepresented minority group.

Now consider two candidates within that minority group: Alice, who is highly qualified, and Bob, who is less qualified. Under the original, unconstrained model, Alice naturally receives a higher score than Bob. However, when the fairness intervention shifts and warps the score distributions to satisfy the global group quota, something unexpected can happen: the scores of individuals within the minority group are shuffled indiscriminately. Suddenly, Bob is ranked ahead of Alice.

This phenomenon is a violation of rational ordering (Lipton et al.). When algorithmic fairness interventions focus exclusively on aggregate group metrics, they often introduce severe in-group unfairness (Yang) and underranking (Gorantla et al.), where highly qualified individuals are penalized or displaced relative to their immediate peers. As Rosenblatt & Witter demonstrate, a fundamental mathematical tension exists between satisfying group equity and maintaining a consistent merit-based hierarchy. Naive optimization frequently results in "rank flipping."

To resolve this tension, a distinct research direction has emerged: intra-group rank preservation. The core intuition is simple. We can alter the distributions to achieve equity between groups, but we must treat the internal, relative ordering of individuals within each group as an unyielding constraint.


The Core Mechanism: Monotonic Score Transformations

How do we guarantee that an intervention never flips the relative order of two individuals in the same group? The most direct mathematical tool is the monotonic transformation.

If the original model assigns a score $s(A)$ to individual $A$ and $s(B)$ to individual $B$, and both belong to the same demographic group, a rank-preserving fairness intervention ensures that:

$$s(A) > s(B) \implies s_{\text{fair}}(A) > s_{\text{fair}}(B)$$

By forcing the mapping from original scores to fair scores to be strictly increasing (monotonic), the internal ordering of risk or merit remains perfectly intact.

Wei et al. and Wei et al. formalized this mechanism through the FairScoreTransformer (FST) framework. Instead of completely re-engineering a model's internal representations, FST acts as a post-processing or pre-processing wrapper. It applies optimal, group-specific monotonic shifts to the predicted probabilities. The authors demonstrate that by framing the problem as a constrained optimization that enforces monotonicity, systems can achieve strict demographic parity or equal opportunity without disrupting the internal consistency of the model's original risk assessments.


Distributional Alignment via Quantile Matching

Another way to understand why rank preservation works is through the lens of quantile matching or optimal transport. Instead of shifting scores by an arbitrary optimization value, what if we map individuals based on their exact percentile standing within their own demographic group?

If Alice is in the 90th percentile of qualifications within her demographic group, a rank-preserving fairness intervention maps her to the 90th percentile value of a shared, baseline "fair" distribution.

Biased Distribution (Group X)          Fair Baseline Distribution
[ Min .... Bob (30%) .... Alice (90%) ] ---> [ Min .... Bob' (30%) .... Alice' (90%) ]
                                 (Quantile Mapping)

This approach is highly effective for data pre-processing and causal fair machine learning:

  • Causal Data Pre-processing: Plecko et al. developed this concept in the fairadapt framework. By identifying an individual's percentile rank within their own group's conditional distribution, the method maps them to the corresponding value in a baseline group. This removes historical group-level disparities while locking the relative standing of individuals in place.
  • Counterfactual Fairness: Similarly, Bothmann et al. and Leininger et al. use quantile preservation to warp real-world data into a fictitious, fair world where protected attributes have zero causal effect on outcomes, ensuring an individual's relative "merit" rank is preserved during the transformation.
  • Optimal Transport in Regression: From a theoretical perspective, Chzhen et al. proved that the mathematically optimal fair predictor for continuous regression tasks is equivalent to the Wasserstein barycenter of the group distributions. By aligning the cumulative distribution functions (CDFs) of different groups, individuals at the same percentile across all groups receive identical adjusted predictions, preserving intra-group ranks by design.

Rank-Constrained Optimization in Lists

When algorithms output an explicit sorted list—such as in search engines, recommendation systems, or shortlists—ranking becomes a zero-sum game. Improving the position of one group necessarily displaces others (Pastor & Bonchi).

To prevent group diversity constraints from shuffling items indiscriminately, systems can inject intra-group fixed-ranking constraints directly into the ranking engine. Vardasbi et al. demonstrate that an optimizer can dynamically adjust the relative exposure or visibility given to different demographic groups while treating the internal, quality-based ordering of items within those groups as a hard, unyielding constraint. This allows platforms to meet equity targets without breaking logical hierarchies or business rules.

However, deterministic rankings present a bottleneck: a single fixed list cannot easily satisfy strict group quotas without causing some individual displacement (Gorantla et al.). To solve this, García-Soriano & Bonchi propose using randomized distributions over rankings. Their "maxmin-fair" framework ensures that within any specific demographic group, a more qualified individual always receives a higher expected utility (e.g., a higher probability of appearing at the top of the list) than their less qualified peers. Crucially, the authors show that preserving internal meritocracy via these randomized distributions results in negligible losses to global ranking quality metrics like Discounted Cumulative Gain (DCG).


Why It Works: Utility, Trust, and Robustness

Prioritizing intra-group rank preservation is not just a philosophical choice; it provides critical practical advantages for machine learning systems:

  1. Preserving Model Utility: Shuffling the relative order of candidates within a group degrades the model's underlying predictive power. By maintaining the original rank order, frameworks like FairOD (Shekhar et al.) can enforce equal flag rates across demographic groups in outlier detection while ensuring that the most anomalous individuals within each group are still prioritized.
  2. Preventing Latent Discrimination: When an algorithm shuffles ranks within a group to meet a constraint, it often does so by exploiting correlated proxies to reconstruct the omitted sensitive data. As Ghili et al. note, preserving the relative ranking of individuals within protected groups prevents this proxy exploitation, making the model more robust and transparent.
  3. Maintaining Stakeholder Trust: If an algorithm ranks a less qualified candidate above a more qualified candidate from the same background, the decision-making process appears arbitrary and unjust.

By framing fairness as a problem of group-level alignment subject to strict intra-group monotonicity, this research direction bridges the gap between systemic equity and individual meritocracy (Brugere et al.). It ensures that the pursuit of a fairer world does not come at the cost of individual consistency.

Key Foundational Work

Does mitigating ML's disparate impact require disparate treatment? (2017) — contributed 2 concept elements to this direction.
Does mitigating ML's impact disparity require treatment disparity? (2018) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 51 papers
  2. Version 1 · 2026-10-04 · 54 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_intra_group_rank_preservation_e28a6da6,
  title = {intra-group rank preservation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intra-group-rank-preservation/rdr_e28a6da6276c45c0a85ae5aa4d9b78d0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_intra_group_rank_preservation_e28a6da6_v2,
  title = {intra-group rank preservation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intra-group-rank-preservation/rdr_e28a6da6276c45c0a85ae5aa4d9b78d0/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake