AI Paper Atlas

Research map / Learning theory and algorithms / Fairness

FAIRER: Fairness as Decision Rationale Alignment

Tianlin Li, Qing Guo, Aishan Liu, Mengnan Du, Zhiming Li, Yang Liu

ICML 2023 · Learning theory and algorithms › Fairness · approach: Fairness auditing and regularization · 4 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add gradient-guided parity alignment (DRAlign) on top of FairRegAdult MLP · Adult-DP at similar AP (~0.770): 41.6% DP improvement over second-best (FairReg(ΔDP, Aug))Improved
increase fairness regularization weight λ from 0 to 0.6FairReg(ΔDP, noAug) on Adult · Adultparity score d_F: 0.624 -> 0.039Improved
increase fairness regularization weight λ from 0 to 0.6FairReg(ΔDP, noAug) on Adult · AdultAP: accuracy drops by a large margin as λ growsHurt
add augmentation (FairMixup) to alignment term (ablation)DRAlign on Credit · Credit-DP and AP: -0.0169 -> -0.0155; AP 0.877 -> 0.881Improved
add augmentation to fairness regularization (FairReg Aug) on CreditFairReg(ΔDP) vs DRAlign on Credit · Credit-DP: FairReg(ΔDP, Aug) better than DRAlign under DPMixed
augmentation on CelebAFairReg(ΔDP) with augmentation · CelebA-DP: noAug slightly higher fairness than AugHurt
oversampling instead of fairness regularizationCelebA wavy hair AlexNet · CelebA-EO vs AP: oversample -0.069 at AP 0.812 vs FairReg(ΔEO) -0.054 at AP 0.817Hurt
remove fairness regularization term, keep only alignment loss (w.o.FairReg-DRAlign) (ablation)Adult MLP · AdultAP and -DP: (0.776, -0.16) -> (0.781, -0.14)Improved
add alignment loss to fairness regularization on the attractive attributeCelebA attractive AlexNet · CelebA attractive-DP and AP: -0.3196 vs -0.3305 at λ=0.1 with AP 0.8956 vs 0.8942Improved
align only last two layers instead of all layersResNet-18 wavy hair · CelebA-DP: -0.3011 vs -0.3063 at λ=0.1 (DRAlign better AP); -0.2481 vs -0.2552 at λ=10.0Mixed
increase over-parameterization (hidden size 10 to 200) (ablation)FairReg(ΔDP, noAug) MLP on Adult · Adultdecision rationale similarity score: c200 reaches only 1.8 at λ=0.6 vs near 3.0 for smaller netsHurt
add gradient-guided parity alignment on COMPAS (EO metric)MLP on COMPAS · COMPAS-EO: -0.044 (DRAlign) vs -0.045 (FairReg ΔEO) at λ=0.05/0.02 comparable APNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy