Feature Attribution and Causal Explanations for Algorithmic Fairness

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2216 papers262 conceptsfrozen 2026-10-04

Standard fairness audits often act like a smoke detector: they can tell you that a fire exists—usually by highlighting a gap in statistical parity or equal opportunity—but they cannot tell you where the fire started or how to put it out. Researchers are moving beyond these aggregate metrics toward fairness attribution, a suite of techniques designed to pinpoint the specific features and causal pathways that drive discriminatory outcomes.

The Diagnostic Gap in Statistical Fairness

Traditional metrics, such as demographic parity, focus on the joint distribution of model predictions and sensitive attributes. However, as Niki Kilbertus, Bernhard Schölkopf, and colleagues argue in Avoiding Discrimination through Causal Reasoning, these metrics are fundamentally limited because they ignore the underlying causal mechanisms. Two models can exhibit identical statistical bias while one is driven by legitimate clinical proxies and the other by systemic prejudice.

To bridge this gap, the field has shifted toward "explainable fairness." The goal is to decompose a model’s total bias into the individual contributions of its input features. If a model is unfair, we need to know if the culprit is a single proxy variable (like "zip code" for race) or a complex interaction of seemingly benign features.

Decomposing Bias with Game Theory

One of the most powerful tools for this decomposition is the adaptation of Shapley values from cooperative game theory. In a standard interpretability context, Shapley values explain how much each feature contributes to a single prediction. In fairness research, authors like Dillon Bowen and Lyle H. Ungar in Generalized SHAP propose treating the fairness metric itself as the value to be distributed.

By defining the "payout" of the game as the difference in average predictions between two groups, practitioners can calculate exactly how much of that disparity is "owned" by each feature. Alexander Asemota and Giles Hooker demonstrate in Targeted Learning for Data Fairness that this allows us to quantify how variables like marital status or education level contribute to violations of demographic parity. This moves fairness from a binary "is it fair?" to a granular "which part is unfair?"

Causal Pathways and Proxy Identification

While attribution tells us which features are involved, causal modeling tells us how they are involved. A feature might be highly predictive and "unfair" simply because it is a proxy for a protected attribute. For example, Yun Wang and colleagues found in Measurement Bias in Machine Learning-Enhanced Opioid Risk Scoring Systems that "number of pharmacies visited" often acts as a proxy for insurance type and race rather than just medical risk.

To formalize this, Bishwamittra Ghosh, Debabrota Basu, and Kuldeep S. Meel introduced Fairness Influence Functions (FIFs) in How Biased is Your Feature?. Their framework, FairXplainer, uses global sensitivity analysis on graphical models to identify not just individual biased features, but intersectional bias—where the combination of two or more features (e.g., gender and age) creates a discriminatory effect that neither feature would produce alone.

Counterfactual Auditing: The "What-If" Test

The most intuitive facet of this research direction is Counterfactual Fairness. Proposed by Matt Kusner, Joshua R. Loftus, and Ricardo Silva in Counterfactual Fairness, this framework asks: "Would the model's decision for this specific individual change if their protected attribute (like race) were different, while all other non-descendant causal factors remained the same?"

This approach works by:

  1. Building a Causal Graph: Mapping how attributes like "Race" influence "Education" and "Income."
  2. Isolating Latent Factors: Identifying "background" variables that are not affected by the protected attribute.
  3. Testing the Flip: Ensuring the model only relies on these "fair" latent factors or paths that are deemed non-discriminatory.

Tools like FlipTest, developed by Emily Black and Matt Fredrikson in FlipTest: Fairness Auditing via Optimal Transport, automate this by finding "counterparts" in different demographic groups. By observing which features must be changed to make a person from Group A look like a person from Group B in the eyes of the model, auditors can unmask the specific proxies the model has learned to use for discrimination.

Why This Works

This research direction works because it transforms fairness from a constraint into a diagnostic. By using game-theoretic attribution (SHAP), influence functions (FairXplainer), and causal counterfactuals, practitioners can move beyond identifying that a model is biased to understanding the generative process of that bias. This allows for targeted interventions—such as removing specific proxy features or augmenting data where the model relies on stereotypical cues—rather than applying blunt, aggregate corrections that often degrade model utility.

Papers in this version

Showing 60 of 216 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 148 papers
  2. Version 1 · 2026-10-04 · 216 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_causal_fairness_ce394e88,
  title = {causal fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-fairness/rdr_ce394e885d264cdda647bfc91d6d6022},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_feature_attribution_and_causal_explanati_ce394e88_v1,
  title = {Feature Attribution and Causal Explanations for Algorithmic Fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/feature-attribution-and-causal-explanations-for-algorithmic-fairness/rdr_ce394e885d264cdda647bfc91d6d6022/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}