Research map / Learning theory and algorithms / Fairness

Fairness-Accuracy Trade-Offs: A Causal Perspective

AAAI 2025 · Learning theory and algorithms › Fairness · approach: Causal and policy fairness in decisions · 11 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.

Analyzes the fairness-accuracy trade-off from a causal perspective by defining path-specific excess loss (PSEL), decomposing total excess loss into path-specific terms with Shapley-style attribution, introducing the causal fairness/utility ratio (CFUR), and proposing a Lagrangian neural learner (CFCL) with binary search over the penalty weight.

Problem

Fairness constraints may reduce predictive utility, and the trade-off between fairness and accuracy has been largely unexplored through a causal lens; it is unclear how much loss each causal pathway from a protected attribute to the prediction incurs when removed.

Approach

Define path-specific excess loss (PSEL) for constraining causal pathways, decompose total excess loss into sums of PSELs, attribute them via Shapley values, and summarize the trade-off per path with the causal fairness/utility ratio (ratio of TV reduction to excess loss), using a Lagrangian neural predictor whose penalty is tuned by a data-driven binary search.

Key results

• On Census 2018, removing the direct effect gives the best causal fairness/utility ratio, and the team concludes substantial discrimination reduction is possible with small excess RMSE.
• On COMPAS, the indirect effect has APSEL around 12% of AUROC while the direct effect is negligible; removing the spurious effect reduces TV by about 3% and indirect by about 5%.
• On UCI Credit, only the indirect effect shows non-negligible APSEL (5% of AUROC) and ATVD (1.5%); direct and spurious effects are not significantly different from 0.

Measured effects reported in the paper

Surprising findings

• On UCI Credit, the authors ultimately keep the unconstrained predictor because mediators are deemed justifiable, despite the indirect effect reducing discrimination.
• Estimated TVD and TVR measures are not statistically different on COMPAS and Census data.

Limitations

• Relies on the standard fairness model assumption of no hidden confounding
• Exponential complexity in the number of causal paths for decompositions
• Experiments limited to three tabular datasets

Experimental setup

Datasets: Census 2018 COMPAS UCI Credit

Benchmarks: Census 2018 COMPAS UCI Credit

Metrics: RMSE 1-AUROC total variation PSEL TVD CFUR

Baselines: unconstrained predictor DE-fair predictor IE-fair predictor

Builds on: standard fairness model path-specific causal fairness natural direct/indirect/spurious effects

Closest papers by meaning

• FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.92)
• Causal Bias Detection in Generative Artifical Intelligence (NeurIPS 2026 · similarity 0.92)
• Mind the Gap: A Causal Perspective on Bias Amplification in Prediction & Decision-Making (NeurIPS 2024 · similarity 0.91)
• Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.90)
• Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.89)
• Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach (ICLR 2024 · similarity 0.89)