Research map / Learning theory and algorithms / Fairness

Procedural Fairness Through Decoupling Objectionable Data Generating Components

ICLR 2024 · Learning theory and algorithms › Fairness · approach: Causal and policy fairness in decisions · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a framework that decouples objectionable data-generating components from neutral ones via reference points and a value instantiation rule, so predictions use only neutral components while meeting Rawlsian procedural fairness requirements.

Problem

Causal fairness notions constrain outcome-level quantities and, when enforced on model parameters, arbitrarily alter neutral components of the data generating process, causing "disguised procedural unfairness" that violates Rawlsian fair equality of opportunity and the difference principle.

Approach

Instead of constraining fitted parameters, fit each local causal module without fairness constraints, then assign fixed reference values to the input nodes of objectionable edges (chosen to benefit the least advantaged) and propagate values downstream via a value instantiation rule, leaving neutral edges untouched.

Key results

• On the simulated linear example, fairness-constrained parameter optimization produces substantial signed relative deviations on neutral edges (e.g., C→Y), showing disguised procedural unfairness.
• On UCI Adult, the framework raises approval rates for the least advantaged female group relative to the unconstrained baseline, whereas Path-Specific Counterfactual Fairness lowers them further for Y=0.
• When objectionable edges are A→Y and M→Y, reference points keep sex unflipped and set marital status to 'married' instead.

Measured effects reported in the paper

Surprising findings

• Fairness constraints imposed only on objectionable components can distort neutral components, producing unintended deviations.
• Reference points did not flip the protected attribute sex; the mitigation instead operated on other inputs such as marital status.

Limitations

• Requires a known causal graph and identification of objectionable edges, which are assumed given.
• Efficient strategies for decoupling are left to future work when additional knowledge about the process is available.
• Scalability and handling of complex (e.g., neural) mechanisms are not fully established.

Experimental setup

Datasets: UCI Adult simulated linear data

Benchmarks: UCI Adult simulated linear example

Metrics: approval rate by group fitted parameter deviation from ground truth path-specific effect

Baselines: Path-Specific Counterfactual Fairness unconstrained optimized baseline No Unresolved Discrimination

Builds on: Path-Specific Counterfactual Fairness Counterfactual Fairness Rawls pure procedural justice

Closest papers by meaning

• Beyond Distributive Fairness in Algorithmic Decision Making: Feature Selection for Procedurally Fair Learning (AAAI 2018 · similarity 0.87)
• LUCID: Exposing Algorithmic Bias through Inverse Design (AAAI 2023 · similarity 0.86)
• Identifying Reasons for Bias: An Argumentation-Based Approach (AAAI 2024 · similarity 0.84)
• Fairness Shields: Safeguarding against Biased Decision Makers (AAAI 2025 · similarity 0.84)
• Position: Political Neutrality in AI Is Impossible — But Here Is How to Approximate It (ICML 2025 · similarity 0.84)
• Fairness on Principal Stratum: A New Perspective on Counterfactual Fairness (ICML 2025 · similarity 0.83)