Research map / Learning theory and algorithms / Kernels and regression theory

Fairness Overfitting in Machine Learning: An Information-Theoretic Perspective

ICML 2025 · Learning theory and algorithms › Kernels and regression theory · approach: Statistical perspectives on generalization · 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 an information-theoretic framework (MI, CMI, f-CMI, e-CMI, ΔL-CMI bounds) for fairness generalization error of demographic parity and equalized odds, using a variance-based Efron–Stein bounding technique instead of sub-Gaussian assumptions.

Problem

Fairness properties achieved during training (DP, EO) may not generalize to unseen data; existing fairness generalization bounds are either restricted to specific algorithms/losses or are KL-based and not computable for deep networks.

Approach

Replace sub-Gaussian/boundedness arguments with an Efron–Stein variance bound for group-dependent fairness losses, yielding MI and CMI-based fairness generalization bounds whose sensitivity term depends on group sizes (shifted harmonic mean), plus a batch-balancing strategy motivated by the bound.

Key results

• Batch balancing reduces test DP on COMPAS by roughly an order of magnitude for some methods (e.g., DiffDP at 2500 samples on gender drops from 0.026 to 0.003).
• The DP CMI bound tracks observed fairness generalization error across ERM and fairness regularizers on COMPAS, showing strong linear correlation in scatter plots.
• Each experimental data point corresponds to 1050 training runs.

Measured effects reported in the paper

Surprising findings

• Some fairness regularizers (e.g., in certain cases) roughly double the fairness generalization error compared to ERM.

Limitations

• Analysis restricted to binary sensitive attributes and binary labels in main results; multiclass extension is only sketched.
• Bounds can be vacuous and the MI-based version is hard to compute for high-dimensional weights, motivating CMI variants.
• Experiments are limited to two tabular datasets (COMPAS and Adult).

Experimental setup

Datasets: COMPAS Adult

Benchmarks: COMPAS Adult

Metrics: demographic parity error equalized odds error fairness generalization error CMI bound tightness

Baselines: ERM HSIC DiffDP PRemover DiffOdd DiffEopp

Builds on: conditional mutual information bounds f-CMI bound loss-difference CMI (ΔL CMI) Efron-Stein inequality Donsker-Varadhan variational representation

Closest papers by meaning

• Fairness via Independence: A General Regularization Framework for Machine Learning (ICLR 2026 · similarity 0.90)
• Understanding Fairness and Prediction Error through Subspace Decomposition and Influence Analysis (NeurIPS 2025 · similarity 0.88)
• fairret: a Framework for Differentiable Fairness Regularization Terms (ICLR 2024 · similarity 0.87)
• Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.87)
• Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.87)
• Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.87)